<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Ondrej Machart on Medium]]></title>
        <description><![CDATA[Stories by Ondrej Machart on Medium]]></description>
        <link>https://medium.com/@ondrej.machart?source=rss-a94f151b844d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*4JWdFzn3SsKtJp_Nu0lx7Q@2x.png</url>
            <title>Stories by Ondrej Machart on Medium</title>
            <link>https://medium.com/@ondrej.machart?source=rss-a94f151b844d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 03:51:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ondrej.machart/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[A Week In The Life Of A Product Strategist In 2026. Fails Included.]]></title>
            <link>https://medium.com/@ondrej.machart/a-week-in-the-life-of-a-product-strategist-in-2026-fails-included-8d0da6a9916e?source=rss-a94f151b844d------2</link>
            <guid isPermaLink="false">https://medium.com/p/8d0da6a9916e</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[work]]></category>
            <dc:creator><![CDATA[Ondrej Machart]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 13:15:25 GMT</pubDate>
            <atom:updated>2026-09-03T19:03:58.233Z</atom:updated>
            <content:encoded><![CDATA[<h2>A Week In The Life Of A Product Strategist in 2026. Fails Included.</h2><p><strong>I’ve been doing product work for a while and the last twelve months rearranged most of it. So I wrote a week down. Five real days at a sports data company – features that shipped, one idea that died on Wednesday, agents that forgot what they were doing, a townhall where the whole company found out what’s coming. Partly a record for myself, partly because I’m curious whether any of it looks familiar to you.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vdIqQwdHjUo5iY5C_TbzVQ@2x.png" /><figcaption>My desktop lately. The meetings and the AI conversations where 60–80% of work happens depending on the day of the week.</figcaption></figure><p>Three months ago I was Head of Product at <a href="http://livesport.eu/">Livesport</a> (known for <a href="http://flashscore.com/">Flashscore</a>, a go-to app for over 100M sport fans around the world) with a team of 4 amazing product managers. Today I’m an individual contributor labeled Product strategist with a team of unlimited overthinking agents who lately use the term “load-bearing” oddly often.</p><p>If you’d asked me two years ago how my job had changed over the previous five, I’d have said not much, really. The company changed. The job didn’t. Then the last twelve months happened, and I decided to write this down before it changes again – which it will, probably by the time you read this. Maybe it’ll be interesting to compare later. Maybe it’s interesting now, because the job has changed for a lot of us, and unevenly. I’d like to know what yours looks like.</p><p><strong>So: a week. Confidential details left out or edited. It’s a week assembled from real days. Every meeting, initiative and task in here happened. At the end I compare it to how this looked two years ago.</strong></p><p><em>Bonus: I can translate this into Czech for my mom and my nieces, who have been asking me what it is I do for a living, really.</em></p><h3>🗓️ Monday</h3><p>I open my laptop. I check the calendar for the agenda, emails, Slack. Then my AI sessions in <a href="https://cmux.com/">Cmux</a> (better than Mac’s terminal). Some of them were running for a few hours over the weekend and have outputs for me I need to review.</p><h4>⚽️ 9–10am: 3D Goals visualisation feature</h4><p>A feature I built with AI in June. Using deep sports data, it turns numbers into a 3D <a href="https://vod.206416.xyz/86016769dc7e1165a162ac196f343f81382a1be8.html">replay of a goal</a>, with stadium roar, fan banners and interesting stats about the shot. Now it’s August and we’re about to roll it out to more leagues. LLM does not get the data (except for fabricated samples) and only builds the visualisation solution and UI.</p><p>Games were played over the weekend and a colleague from the Data team sent me new feedback. Some goals missing the xG statistic. Some assists landing the ball at the feet of the scorer, then teleporting it to his head instead of landing there directly. LLM gets the feedback, investigates, identifies the root cause, makes the fix and lets me review the changes before deploying to production. I get my coffee while it runs and jot down a few ideas for improvements to discuss later. Meantime, deployed.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r4Tm1TRi5FQ1nPsvFyF4OA@2x.png" /><figcaption>3D Goal Visualisation Lab — Completely vibe-coded, generating videos live during the match.</figcaption></figure><p>It lives on its own deployment space, separated from the rest of the product, so I can deploy independently. It lands well. I relay the good news to my colleague.</p><p>The feature is promising. The younger audience especially has given us great feedback, so it’s worth chasing. But I know I’ve been mostly fixing minor issues or adding nice-to-have touches lately. It will soon scale, get evaluated, then move to BAU. And I’ll get my hands dirty with another challenge.</p><p>It’s not a one man show. I’m a non-technical product guy – so I benefit greatly from sitting next to our Head of AI – Marek – who lends me his mind and time when I need it – to make the solution a real thing, not just a half-baked vibe-coded prototype. I always feel like a single cell organism when I ask him things. Thankfully he’s not just smart but also endlessly helpful. Our Security team provides the tool for vulnerability scans. Our Infra team provides the reliable and scalable deployment space. I still feel nervous. It’s out there, soon used by possibly millions of users. How will this play out?</p><h4>👕 10–11am: Leadership team sync</h4><p>Our regular meeting. Honestly I used to dread those. I had easily 8–9 meetings every day and this one always felt like a waterfall of information. I noticed the volume of meetings is much smaller now. I started to enjoy this sync perhaps because it’s just nice to meet the team. We still share the updates, raise roadblockers, receive new assignments. But somehow I’m enjoying the fact that it’s just people, no agents.</p><p>I’m asked to pick up a few tasks, provide updates on the progress of the ones I work on. We move on.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rkXVDb7D_H5Vd3hBhRosiQ.jpeg" /><figcaption>Livesport offices in Prague. People enjoy coming here. Me too.</figcaption></figure><h4>🤖 11am-12pm: Creating product agents and context</h4><p>One of the big ambitions we have at Livesport is to create agentic workflows that can help the product team with some – if not most – of their tasks – the agents should help prepare for an important meeting, ideate solutions to a problem, prototype ideas, stress test a business case, write a YouTrack ticket, find relevant existing user insights, etc. For this to work, these AI agents need a good definition of who they are, who they are not, what tools they can use and what context to draw from.</p><p><strong>Context is the tricky part. It’s the messy reality of scattered information PMs are used to just knowing after years of working in the company.</strong> Vision and mission engraved on the walls. Objectives. The proclaimed ones and the real ones. Products. User personas. Business model. Product principles, the do’s and the no-nos. Design system. Brand. Tone of voice. Analytical data. Results of AB tests and which ones we overruled. Roadmap. Structure of the YouTrack ticket. Some of it is stable. Some of it changes weekly. Some of it is words, some numbers, some visual. Something lives on Confluence, something in Tableau, some stuff is a markdown document derived from emails.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/563/1*tvdJn4-WGTKqL009oFjeWg@2x.png" /><figcaption>One of my agents, the Validator.</figcaption></figure><p>A colleague from Engineering shared an MCP he created that gives my agents access to over 3k test scenarios and documentation of existing features. I asked my LLM to integrate it into the context and agents. The Critic, Specifier and Prototyper agents will benefit from it immensely. I give them a task where they can use the MCP. They are ecstatic. They always are. I watch them act on it.</p><p>Then I ask another Engineering colleague, Lukas, to schedule a walkthrough of their AI Factory framework. Our context and agents could be moved there and share one logic with the technical ones. The mess could finally fit into neat boxes and shelves. Hopefully. It’s all so new I learnt to just go with the flow and see where it takes us.</p><h4>🌮 Lunch</h4><p>A teammate invites me to join for lunch. I used to skip lunches, or do a quick one to get back to the unfinished work. Not very healthy. Physically and socially. My wife hates this workaholic habit of mine. I love being in a flow and I struggle to pull my head out of it. This AI era does not help. But I’m working on it. So I join, this time.</p><h4>🎾 12.30–1.30pm: Tennis feature improvement</h4><p>A minor but nice improvement idea for our tennis match detail page landed in my email a few days ago. We discussed it, I prototyped a few solutions. To converge, I discussed it with colleagues, PMs and a UX designer. He’s impressed by the fact I let AI pitch the design. By impressed I mean composed enough to not act hurt and ego-less enough to instead be helpful and suggest meaningful changes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iFlsSgGsQ8WR1-HvI3MBkQ.png" /><figcaption>A subtle change. The last point was never rendered in the Game. We’re adding it. Mobi AI agent is, that is.</figcaption></figure><p>Now I’m handing it over to the Engineering team’s AI agents, since it’s a fairly simple task. I give my AI the details – the why, the winning design. It reaches through my Slack to the web and apps AI agents for technical details. Then it drafts the YouTrack ticket. I review it, approve it. The agents pick it up. They come back with a plan and start implementation. I send a few ideas for improvement of the agents to the Engineering colleagues who made them.</p><p><em>You might wonder why a minor thing like this is on the backlog of someone labeled as Product strategist. Two reasons. One is to stay in touch with the product. I don’t want to fly high in the skies talking about some abstract strategy. Second, to shape the agentic workflows, I need to eat my own dog food and go through the hoops.</em></p><h4>🙌 1.30–2pm: Regular 1–1 with Head of UX</h4><p>Every other week I meet with my peer, Matej. We discuss all sorts of things. Life. Work. AI. Kids.</p><p>He shows me an agent his UX team has been working on. It takes a PM’s requirements and turns them into Figma components. A designer must still review it and fix it occasionally but it’s exciting.</p><p>To design new solutions is fun. To turn a design into a component is fun mostly just for OCD designers. So I see this as a freedom for UX designers to do more of what they actually enjoy. I talk from my own experience, I was a UX designer for 10 years before becoming a PM.</p><h4>🌪️ 2–3pm: Multi-tasking window</h4><p>I reply to Slack messages. Check emails. Check new calendar invites for the rest of the week. The tennis improvement got built and the ticket was reassigned to me. I check the testlink. I consult with LLM, then we agree on change requests. He writes them as a comment to the ticket. AI agents virtually nod and take another round of implementation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/929/1*vWbl56oGwxCtg2B4XEgzLw@2x.png" /><figcaption>YouTrack ticket picked up by the Web team’s AI agent Reeman. And a conversation between him and LLM, commenting on my behalf.</figcaption></figure><p>I create a ticket for the Design team agent Matej showed me earlier because I can’t wait to give it a try. Then I check the agenda of the next meeting and add a few points to discuss. I turn over to Marek, the Head of AI, and ask him if an idea I have about the goal visualisation feature is technically feasible. This time he’s pessimistic. Feasible, but it would work only on half of the devices our users typically use. I drop it. I get another coffee.</p><p>Then I review the output of my agents who tested the MCP from earlier that day. It worked wonders. It also added a funny confusion to the agents. They iterated so many times they compacted their memory, forgot the original task, started following a completely different assignment they came up with and dropped the MCP entirely. I guess we’ll need to iterate.</p><h4>👕 3-4pm: Stream leaders sync</h4><p>At Product we’re organized into value streams. Content, Monetisation, etc. We meet every other week to discuss things – who’s presenting what to the owners and management. Who’s building something that could overlap with another stream’s agenda. What needs to be escalated to our CPO. Who hit the LLM limit today.</p><h4>👋🏻 Afternoon wrap up</h4><p>I reply to Slack, check tomorrow’s calendar, stop by TP (our CPO). He shows me a cool football video he made with SeeDance AI. We agree the rate of change is insane. We love it. And are unsettled by it, at the same time. Just roll with it, he says.</p><p>I roll. To my car, to head home. After my precious family time and before going to sleep, I give my agents a few instructions. For the 3D goals, to explore the API documentation and check whether we could add an interesting statistic before/after swing after the goal. And for tomorrow, to start collecting user feedback and reviews on a product we might acquire one day. Good night. For me, at least. Btw – do agents dream? And if so, does that eat tokens?</p><h3>️️️🗓️ Tuesday</h3><p>I won’t be so detailed about the rest of the week. It would turn the article into a chronicle, which is not the goal. Instead, here’s the day in a few bullet points to paint the picture.</p><ul><li>🤖 Anthropic and OpenAI both announced a limit increase. What’s going on this time?</li><li>⚽️ The idea for the 3D goal viz is feasible as per my agent. I ask him to prototype it.</li><li>🕵🏼‍♀️ The feedback analysis agent collected over 130k user mentions of the product. Neat. I ask him to tag each with themes for later analysis.</li><li>👕 Team meeting with TP. Updates from yesterday’s management meeting. Roadmap check. Then mostly admin and HR stuff and some new tasks. Company event next week. Don’t get drunk, you represent the team. It’s ok. We’re high from the LLM limit increase announced earlier this morning.</li><li>️🎙️ Meeting with Martin, our Chief Marketing Officer. I interview him about his views on the 2027. It’s part of my assignment to interview the c-suite and prepare themes for their strategic workshop in September. I hand over my notes to LLM afterwards. It keeps track of it all and suggests new questions for the next interview. I like some, dismiss the rest.</li><li>👕 A feature proposal walkthrough meeting with TP. Tough love feedback, as always. Lots of probing both into nitty gritty details and the higher level whys behind. Steering, solutioning, some conclusions, some new questions to find answers to. Next round in a few days.</li><li>👨🏻‍🏫 A session with my mentor. Happens irregularly but always opens my eyes to a new perspective. Can’t stress enough how impactful it is when you find the right mentor for you.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UA6Vm8xddASLL3S4ms4sWA@2x.png" /><figcaption>Competition analysis microsite built with LLM. Offers interactive glimpse into the collected data.</figcaption></figure><h3>🗓️ Wednesday</h3><ul><li>⚽️ The 3D goal idea prototype looks good. But the data coverage for the statistic swing is not optimal. I ask Radek, the API owner, if it can be boosted. It can’t at the moment. The idea gets parked. Knowing Radek, he might come back in a week he found a way.</li><li>🇪🇸 A meeting with TP about agenda for my upcoming business trip to Malaga, where our sister company BeSoccer is located. I’ll help them with AI adoption and in return learn about how they approach product work. Can’t wait to meet the team. And go to the beach every evening.</li><li>🕵🏼‍♀️ The reviews analysis repo got new inputs from Martin, a colleague from BI. He collected and annotated over 130 screenshots and added a feature gap analysis. I ask LLM to draft an internal microsite to start surfacing the insights.</li><li>🎙️ Another interview with one of our C-level managers. Kavic, our CTO. I feel like I asked him less than half of the questions I had. I did not manage the time well and under-moderated the discussion. I’ll send him more questions over Slack.</li><li>🆎 A meeting with Prokop, our Growth engineer, to align on an AB test of a new widget in match detail. Part of it will be vibe coded by Honza, a PM colleague from Monetisation stream.</li><li>🎾 The tennis improvement built by the AI agent passed testing on web. I accept it. It will launch next week.</li><li>📊 I kick off a new project – an engagement composite score to give the word <em>engagement</em> an actual value we can compare over time, across the portfolio. I start by using my suite of agents to source the data from the Tabelau. This will take some time to keep them rooted in validated, correct data.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sgrMZN_ZEPB6dRAWDknnEQ@2x.png" /><figcaption>My Engagement score report (edited). Did not land well on the first try. But it makes sense to iterate.</figcaption></figure><h3>🗓️ Thursday</h3><ul><li>👕 Sprint review of the mobile apps team. What goes into the next release, what’s still in progress. The tennis improvement did not make it there unlike on web. Turns out the agent ignored app components and just hard coded stuff willy-nilly. He needs that design agent counterpart Matej showed me. Don’t we all?</li><li>💡 Feature solution walkthrough with owners. Lots of new perspectives. Some constructive pushbacks. Valuable insights into how they think about the product. Iterations on the spot, using LLM.</li><li>🤖 An AI consultation window I hold weekly for my teammates. Part of my role is to help the team upskill in AI. Last week nobody needed anything, everyone was busy. I felt a bit stupid. But this time two people showed up. One asked for advice. The other for feedback. I still fight imposter syndrome. I have been fighting it most of my career. Especially now. I am no AI expert. Just slightly ahead of my colleagues, for now. It feels awkward to have this consultation window. I still learn as much as others do. I’ll keep the window until I figure out a better way.</li><li>⚽️ Preparation of a slide for TP, for his upcoming monthly business review with the board. One of the few presentations still made with Google Slides.</li><li>🕵🏼‍♀️ Molding and sculpting the microsite with the competitor app reviews analysis, through AI. Asking LLM to review hypotheses I prepared for Carol, our senior researcher, to validate with users.</li><li>📊 I showed the engagement composite score draft to a few people. They think it’s cute and that nobody will understand it. Keep it simple. Show the metrics first, don’t start with an abstract number you have to explain like a bad joke. Start over.</li><li>⚽️ Meeting with Martina, our lead researcher. She shows me the results of her second quantitative study on the 3D goals. The first one was so positive. But the new 2D variant that is faster to render and hides some of the weirdness the 3D animations invite landed worse with users. Wrong assumptions. We’re keeping the 3D. We just have to make it better.</li><li>🤖 Monthly AI jam, a session I hold for the whole product team where we exchange tips on how we use AI, skills, agents, projects. Petra, our DesignOps team leader, shares updates on her shared repo of baby Flashscore – anyone can grab it and prototype into it. Love it.</li></ul><h3>🗓️ Friday</h3><ul><li>📝 Time to write our weekly product dept status update that gets sent around the company.</li><li>🤖 I try to close all the loose ends. The win probability prototype — case closed, code cleaned up. The competitor analysis microsite goes around internally, it looks pretty good. The engagement score iteration — finally making more sense at the first sight.</li><li>📖 I read through material from TP, our CPO, on the next big thing. A key page evolution plan and concepts. I love big new challenges with lots of open space to explore.</li><li>💬 Then he drops me a Slack message. “Have you read about the Forward deployed product manager role?” With a link to an <a href="https://www.secondtalent.com/occupations/forward-deployed-product-manager/">article</a>. My role changed three months ago and here comes the next one. I guess changes come and go like waves. Roll with it, Ondra.</li><li>🏢 CEO Townhall meeting where he speaks to the whole company. This time with the CPO and CTO about the plans with AI factories. These should not take people’s jobs away. They should shift people’s value from the details execution to the orchestration, architecture, principles, context, evaluation. Applies not just to Engineering work but Product, too. You can sense the energy in the room. The times they are a changing. Some of us are excited, some uneasy, most are both, or oscillating from day to day. Every company goes through unknown waters. We’re heading there, and not everybody is up for the ride. The cultural challenge of the changes coming is enormous. Everybody on the market feels the tectonic plates are shifting. We in the office feel it too. The way we approach this, how we make it possible for people to use the shift to grow together, is critical.</li></ul><p>Discussing the Townhall with teammates on the way back to our space. Interesting week. Agents still crunching numbers, with the amphetamine app preventing sleep. But I go home. We go camping. Hiking. Live the real, offline life. It’s refreshing. I love the job but the screen time is just too much. It was never this bad. We’ll need to fight it. Keep things human. Despite what’s coming. It’s still just a tool and we should treat it that way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zi5hNhuDwboEQN58BpWofg.jpeg" /><figcaption>Off to English nature with my son Filipek and wife Iveta.</figcaption></figure><h3>🌱 How is this different from two years ago?</h3><p>The substance hasn’t changed all that much. We still build the product. Fight that User – Business – Tech triangle of balance. The Tech part is changing a lot. Not yet the big pieces of it. But AI factories are coming. PM and UX designer roles blur and overlap. Builders at heart emerge.</p><p>But we still have stakeholders to cater to. Users to understand. Data to measure and interpret. Decisions to make. Responsibility to face. Deadlines, bottlenecks. The shift and move. The ways of working changed. I have fewer meetings. I always have 3–5 LLM Code sessions running. I channel most things through them. My attention and value added in some tasks moved from the middle (the execution) to the start (inputs, context, requirements) and the end (eval, feedback, iterations).</p><blockquote>The substance hasn’t changed all that much. We still build the product. Fight that user — business — tech triangle of balance. The tech part is changing a lot. Not yet the big pieces of it. But AI factories are coming. PM and UX designer roles blur and overlap. Builders at heart emerge.</blockquote><p>It’s honestly both fun and exhausting. Fun because it is so creative. I get to build. As a kid I’d spend hours with LEGO. It was a freedom of building that filled me with joy. The new role brought back those memories and feelings. But it is also exhausting because I switch between 5–7 contexts every day. LLM dumps 500 words, 3 prototypes to test, 5 decisions to make with 18 options to choose from, and it goes in rounds that spin every few minutes.</p><p>But I see my value has lifted. I am a few prompts away from kickstarting something meaningful, always. I just have to learn to use the tools effectively – both for my sake and the business needs.</p><p>Now off to the campground.</p><p><strong>And how’s your week been lately?</strong></p><blockquote>🫶🏼 Note 1 – The distribution of changes is always uneven. Your role, company, product, context differs. I am far from telling you how it should change. The way we do things sometimes works, sometimes not. We’re learning. Your path is yours and you control it. Do what’s best for you.</blockquote><blockquote>✍🏼 Note 2 – I wrote most of my articles with the help of AI. The thoughts were mine but I let it construct the structure and the language, since English is my second. It came out polished but somehow without my soul or personality. So this one I wrote purely myself. It’s not perfect. But I enjoyed it. And I hope you did too.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8d0da6a9916e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Building Gets Easier, Decisions Get Harder]]></title>
            <link>https://medium.productcoalition.com/you-can-ship-anything-now-can-you-kill-anything-a12ca5636c16?source=rss-a94f151b844d------2</link>
            <guid isPermaLink="false">https://medium.com/p/a12ca5636c16</guid>
            <category><![CDATA[product-design]]></category>
            <category><![CDATA[product-development]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ondrej Machart]]></dc:creator>
            <pubDate>Fri, 14 Aug 2026 01:54:21 GMT</pubDate>
            <atom:updated>2026-11-02T19:50:28.075Z</atom:updated>
            <content:encoded><![CDATA[<h4>How does the role of a product manager change as AI makes building software faster and easier?</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DiNPyAnuYqFloOPUCfiyiQ.png" /><figcaption>The Art of Momiage.</figcaption></figure><blockquote>In Japanese gardening, the art of Momiage reminds us that a tree doesn’t always grow healthier by adding more. Sometimes, carefully removing what is no longer needed allows the tree to thrive.</blockquote><blockquote>I think a similar idea applies to product management, especially as AI makes it easier and faster to build software. We may soon find that our biggest challenge isn’t how much we can create, but how well we decide what deserves to stay. A recent experience at Flashscore brought this possibility into focus for me.</blockquote><p>A while ago, while working at Flashscore, a sports platform that provides live scores, match results, and statistics, a request to improve how tennis fans follow individual points during a match landed on my desk.</p><p>As usual, the team discussed what needed to change, why it mattered, and how the improvement should work. Once we agreed on the approach, the change was built using our AI-assisted development process.</p><p>What surprised me was how quickly it happened. For the first time, we had spent more time discussing what to build than actually building it. But the speed wasn’t the only interesting part.</p><p>I watched AI agents participate in our development workflow, responding to comments in YouTrack, our project management system, generating a link to test the change, and even questioning a proposed revision.</p><p>I wasn’t communicating with these agents directly. My Claude AI assistant was handling those conversations, using its knowledge of Livesport, the company behind Flashscore, to communicate within our existing workflow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KHhtxaG7cA02SDGjshArpw.png" /><figcaption>Snapshot of Youtrack conversation with Reeman, Livesport’s web AI agent.</figcaption></figure><p>It was a small improvement to one part of our product, but the experience made me think about something much bigger. For years, product managers have worked within the limits of how much software their teams could build. Engineering time was valuable, so choosing which ideas deserved attention was an important part of the job.</p><p>But what happens when AI makes building software much faster and easier? If teams can build more features in less time, the challenge may no longer be deciding only what to build. It may increasingly involve figuring out which ideas actually work, which features improve the customer experience, and which ones should be removed. Based on what we’re beginning to experience at Livesport, that’s the change I want to explore.</p><h3>The Prioritisation Paradox</h3><p>For years, one of the biggest challenges in product management was the time and resources needed to build and launch new features.</p><p>There was never a shortage of ideas. The problem was that development teams had limited time and capacity, so product managers had to carefully decide which ideas were worth pursuing, often without knowing for certain whether they would succeed.</p><p>Most ideas never made it into the final product, which wasn’t necessarily a bad thing. These limitations forced teams to question ideas early and abandon those that seemed unlikely to deliver value.</p><p>But there was another side to this. Because developing and launching a feature often took months, teams became heavily invested in their decisions before they could see how customers actually responded. And when the results were disappointing, it wasn’t always easy to admit that an idea hadn’t worked.</p><p>Sometimes teams didn’t measure the results carefully enough. Other times, they interpreted the data in ways that justified keeping the feature. It wasn’t necessarily intentional. It’s human nature to defend something you’ve already invested considerable time and effort into.</p><p>So while limited development resources helped teams avoid pursuing weak ideas, they also made it harder to walk away from features that had already consumed significant time and effort.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vAAPE_cWNuxSLL3lHJM4uA.png" /><figcaption>Japanese garden water basin.</figcaption></figure><h3>The bottleneck doesn’t disappear. It simply moves.</h3><blockquote>Think of the Japanese garden water basin pictured above. Water flows steadily through a bamboo pipe into the basin, but the basin can only hold so much before it begins to overflow. Making the water flow faster doesn’t increase the basin’s capacity.</blockquote><blockquote>Likewise, AI may allow product teams to build and launch features much faster, but that doesn’t automatically increase their ability to test those features, understand how customers respond, or determine which ones are worth keeping. The challenge hasn’t disappeared. It has shifted from how quickly we can build something to how effectively we can evaluate what we’ve built.</blockquote><p>Think about it. If building a feature becomes faster and less expensive, why not let customers try it and use their feedback and actual behavior to determine whether it should stay?</p><p>This approach isn’t entirely new. Companies have long used A/B testing to compare different versions of a feature and see which performs better. Booking.com, for example, has reportedly conducted more than a thousand experiments at the same time, using the results to guide product decisions. What AI could change is how widely and frequently teams can experiment.</p><p>But faster development doesn’t eliminate the need to make difficult decisions. It creates a different set of challenges. Teams will need to measure results, understand customer needs, maintain a consistent product experience, and decide when it’s no longer delivering value.</p><p>In other words,<em> building and launching features may become easier, but deciding which ones deserve to remain could become much harder. </em>All of this raises some interesting questions about where product management is heading, what might change, and what will matter most as AI becomes a bigger part of how we build and improve products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OhGUWAYP9XTNU6LRs3ggiQ.png" /><figcaption>A Masu box.</figcaption></figure><h3>Measurement becomes the new bottleneck</h3><blockquote>The wooden box pictured above is called a masu, a traditional Japanese container used to measure rice. Its size determines how much it can hold, regardless of how much rice is available.</blockquote><blockquote>Similarly, AI may allow product teams to build and launch many features at once, but measuring their individual impact is another matter. When too many changes are introduced simultaneously, it becomes difficult to determine which ones are actually improving the product.</blockquote><blockquote>This creates a new challenge for product managers: making sure the pace of development doesn’t exceed their ability to measure results and learn from them.</blockquote><p>This means product teams will need to be more thoughtful about how they introduce changes. Rather than releasing everything at once, they should create opportunities to measure the impact of each feature. They’ll also need to distinguish between the initial excitement of something new and the lasting value it brings to customers. Some features attract attention because they’re new, while others may take time for customers to understand and appreciate.</p><p>I expect product managers to become much more involved in analytics, using AI to help them collect, interpret, and understand the data. Not because AI makes these activities easier, but because **decisions about whether to keep or remove a feature should be supported by evidence rather than personal opinions.**</p><p>But there’s another limitation that AI cannot easily solve. Even if a team can build a hundred features, it doesn’t necessarily have enough customers to test all of them properly. Each experiment needs sufficient participation to produce meaningful results.</p><p>Consider Booking.com, which has been running large numbers of experiments simultaneously for years. Its enormous customer base gives it the ability to detect even small changes in customer behavior. A company with hundreds of millions of users can potentially identify a 0.1% improvement in conversions within a day. Most product teams don’t have that kind of traffic or the same opportunity to experiment at such a scale.</p><p>This creates another important decision for product managers. Before deciding which features to remove, they may first need to decide which experiments are worth running.</p><p>If a team has a limited number of users, running too many experiments at once can make it difficult to draw reliable conclusions. Choosing which experiments deserve the available audience becomes another form of prioritization. In the past, limited development capacity often made these decisions for us. Now, teams may need to make them more deliberately.</p><p>But what happens when a product doesn’t have enough users to support frequent experiments?</p><p>In that situation, launching every idea just to see what happens isn’t necessarily the best approach. It could leave the product crowded with features that haven’t been properly evaluated.</p><p>Instead, teams can take advantage of faster development much earlier in the process. Rather than spending weeks discussing a single concept, they could use AI to build several working prototypes, compare different approaches, and gather feedback before deciding what deserves a full release.</p><p>For smaller product teams, the greatest benefit of AI may not be the ability to launch more features. It may be the opportunity to explore more ideas, test assumptions earlier, and make better decisions before investing in a full launch.</p><p>When teams introduce a new feature, they should agree in advance on what success looks like, how they’ll measure it, and what results would justify changing or removing it. And once those expectations are established, they need to be taken seriously. <em>There’s little value in setting measures of success if we’re unwilling to act when the results don’t meet them.</em></p><p>Ultimately, faster development makes measurement and evaluation even more important. The goal isn’t just to build more. It’s to learn what works, what doesn’t, and use that knowledge to make the product better.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GTNp1f-ojFvRLu30bnxiEw.png" /><figcaption>An expanded Japanese Zen rock garden.</figcaption></figure><h3>Coherence becomes the design job</h3><blockquote>The Japanese Zen garden pictured above is carefully arranged to create a sense of order and balance. Its sand can be raked into new patterns, while the stones remain in place, preserving the garden’s familiar structure even as its appearance changes.</blockquote><blockquote>As AI makes it easier to introduce new features, product teams face a similar challenge: allowing their products to evolve without disrupting the familiar experience customers have come to understand and rely on. But how much change are customers willing to accept before they become frustrated?</blockquote><p>People aren’t always comfortable with change, especially when it affects something they’ve grown accustomed to using. But we’ve also shown that we can adapt. We regularly update our apps, adjust to changes in social media feeds, and learn new ways of interacting with software that continues to evolve.</p><blockquote>The bigger problem isn’t necessarily how often a product changes. It’s whether those changes continue to make sense to the people using it.</blockquote><p>Imagine returning to an app you’ve used for years, only to discover that a familiar feature now works differently, or that something you once found easily has been moved without an obvious reason. Even if the change was intended to improve the experience, it can create confusion when it doesn’t fit with how customers already understand and use the product.</p><p>Over time, these small inconsistencies can accumulate. An old feature remains even though a newer one serves the same purpose. A setting that was once useful is rarely touched. Different parts of the product begin to behave in different ways, making the overall experience harder to understand. This is why I believe maintaining a consistent and understandable product experience will become an increasingly important responsibility.</p><p>As AI makes it easier to introduce new features, someone still needs to make sure those additions fit naturally into the product and that unnecessary complexity doesn’t build up along the way. And this could change how several roles within product teams operate.</p><h4>User Research</h4><p>User research will remain essential, and I suspect it may become even more important as development moves faster.</p><p>Analytics can tell us what customers are doing, which features they use, and where they encounter difficulties. But numbers alone don’t always explain why people behave the way they do.</p><p>Traditionally, much of our research has focused on understanding customer needs before a feature is developed. That work will still matter, but faster development creates more opportunities to learn from customers after they’ve experienced a working product.</p><p>Instead of relying mainly on what people say they might do with a proposed feature, we can observe how they actually use it, speak with them about their experience, and combine those conversations with the data we’ve collected. I believe this will give product teams a stronger foundation for deciding what to improve, what to change, and what may no longer be necessary.</p><h4>Designers and DesignOps</h4><p>I also expect the responsibilities of designers and DesignOps teams to evolve. As AI makes it possible to turn ideas into working prototypes much faster, some designers may become more directly involved in building those prototypes themselves. Others may spend less time producing individual screens and more time making sure the overall product experience remains consistent.</p><p>This distinction matters because a product can have attractive individual features and still feel confusing when those features don’t work well together. Designers will increasingly need to help establish the principles, patterns, and systems that allow teams to move quickly without creating unnecessary inconsistencies. We’re already seeing this happen at Livesport.</p><p>Our DesignOps team previously focused on maintaining our design system in Figma, the tool we use to design and organize product interfaces. As AI became more involved in development, the team adapted that system so AI tools could work with it through Figma’s Model Context Protocol (MCP) integration.</p><p>They then created what we call <em>baby Flashscore</em>, a working version of our product’s underlying structure that product managers can use to build and test prototypes.</p><p>This means our PMs can now develop ideas within the same design framework used by the actual product, rather than working from separate mockups that may not fully reflect how everything fits together.</p><p>Our UX designers have also changed how they work. Previously, a product manager might describe an idea, and a designer would create the screens in Figma to show how it could look and function.</p><p>Now, designers and product managers can work together on the same functioning prototype. A PM can connect that prototype to our sports data systems and demonstrate how it handles real matches, live statistics, and unusual situations, such as a football match ending without a single shot on target or being abandoned after twelve minutes.</p><p>Because the prototype uses real data, the team can explore these situations together and make changes during discussions rather than repeatedly passing designs back and forth. For me, this represents an important change in how product teams collaborate.</p><blockquote>The product manager and designer are no longer handing work from one person to another. They’re working together to shape the experience as it develops.</blockquote><p>The designer continues to represent the customer’s perspective, brings ideas the product manager may not have considered, and helps ensure that new features fit naturally into the product. And as the pace of development increases, I believe that responsibility for maintaining a consistent, and understandable experience will only become more valuable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6z4MUu8XWFNmV5XoeAHOtA.png" /><figcaption>Ikebana</figcaption></figure><h3>Product taste becomes visible, but with a caveat</h3><blockquote>The Japanese art of Ikebana, pictured above, involves carefully arranging flowers and branches to create a sense of beauty, balance, and harmony. But what makes the arrangement meaningful isn’t just what the artist chooses to include. It’s also what they deliberately leave out.</blockquote><blockquote>Product managers face a similar challenge. Good judgment isn’t always reflected in the features they introduce or the results those features achieve. Sometimes, their best decisions involve recognizing which ideas aren’t worth pursuing or which features would add unnecessary complexity. Those decisions may be less visible, but they can be just as important to the quality of the finished product.</blockquote><p>For years, it wasn’t always easy to recognize a product manager’s ability to make good product decisions. Ideas often passed through several layers of discussion, approval, research, and development before reaching customers. Along the way, the original idea could change considerably.</p><p>A senior executive, sometimes referred to as a HiPPO (Highest Paid Person’s Opinion), might reject a promising idea because another project was considered more important. Research findings, design revisions, and competing priorities could also reshape an idea until little of the original thinking remained.</p><p>As AI makes it easier and faster to turn ideas into working products, product managers may have more opportunities to demonstrate the quality of their judgment. Instead of spending months defending an idea, they can test it, gather feedback, and see whether it delivers the expected results.</p><p>This doesn’t mean every idea has to work perfectly the first time. Good product judgment is often about identifying the right direction, learning from what happens, and making improvements until the desired outcome is achieved. Over time, product managers who consistently make sound decisions may become easier to recognize.</p><blockquote>But there’s an important limitation. Not every good product decision can be measured through numbers. Analytics can help us understand whether a feature improves engagement, increases conversions, or achieves another business goal. But they don’t always capture the value of keeping a product simple, maintaining a consistent customer experience, or choosing not to introduce a feature in the first place. If we judge product managers only by the results of what they build, we risk overlooking some of their most valuable decisions.</blockquote><p>I don’t think the answer is to introduce more measurements. Instead, we should pay greater attention to the thinking behind each decision and whether the product manager’s expectations proved accurate.</p><p>Before launching a feature, a PM should clearly explain what they expect it to accomplish, how much improvement they anticipate, and what results would suggest that the feature needs to be changed or removed.</p><p>The performance of a single feature tells us something about that feature. But looking at a series of decisions over time can tell us much more about the product manager’s judgment, including how accurately they anticipate outcomes and respond when things don’t go as expected. And we shouldn’t overlook the value of deciding to remove something.</p><p>Most organizations celebrate the features they launch and the improvements they introduce. Far less attention is given to the decisions that prevent unnecessary complexity or remove features that no longer serve customers.</p><p>If we only recognize product managers for what they build and keep, we encourage a culture where adding more is rewarded, even when doing less might produce a better experience.</p><p>A product manager who recognizes that a newly launched feature isn’t working and recommends removing it after two weeks may have demonstrated better judgment than someone who continues investing in a feature because it has already been built.</p><blockquote>Ultimately, good product taste isn’t just about knowing what to create. It’s also about recognizing what to improve, what to leave alone, and when something no longer deserves a place in the product.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AAWDZs8irSbK8pjyxfEf-g.png" /><figcaption>Ise Jingu shrine.</figcaption></figure><h3>Killing becomes a skill PMs and companies have to build</h3><blockquote>Every twenty years, Japan’s Ise Jingu shrine undergoes a remarkable process in which its important buildings are carefully reconstructed as part of a centuries-old tradition. Rather than waiting for the structures to deteriorate, rebuilding is an intentional part of preserving them. Product management could benefit from a similar approach.</blockquote><blockquote>Keeping a product useful requires more than introducing new features. It also means regularly reviewing what already exists, simplifying what has become unnecessarily complicated, and removing what no longer serves a purpose. As AI speeds up development, this becomes even more important. Organizations need to create a culture where these decisions are not only expected and supported, but also recognized as valuable contributions to improving the product.</blockquote><p>Everything we’ve discussed so far assumes that product teams are willing and able to remove features when they’re no longer delivering value. But in practice, many organizations find it much easier to introduce something new than to take something away.</p><p>One reason is the amount of time, effort, and resources already invested in a feature. Before launch, a company may be reluctant to abandon an idea because of what it has already spent developing it. But after launch, removing that feature becomes even more complicated.</p><p>Customers may have started using it. Support teams may have created documentation explaining how it works. Sales teams may have introduced it to clients, and employees may have celebrated its release as an important achievement. By that point, the feature has become part of the product and the organization’s commitments.</p><blockquote>AI may make building software faster and less expensive, but removing something that already exists can still be surprisingly difficult.</blockquote><p>And the difficulty isn’t always technical. Removing a feature can affect customers, colleagues, and departments that have come to depend on it. It can also challenge decisions that people have spent months defending. We’ve removed a few products at Livesport, and those decisions were never easy. I don’t believe the biggest obstacle is that product managers lack the ability to recognize when something should be removed. The problem often lies in how organizations encourage and reward certain behaviors.</p><p>Product launches are celebrated. New features appear in performance reviews, team accomplishments, and company announcements. Employees can point to what they’ve built as evidence of their contribution. But removing an unnecessary feature rarely receives the same recognition, even when it makes the product better. Over time, this creates an environment where adding more feels like progress, while removing something can feel like admitting failure.</p><p>As AI makes development faster, I expect we’ll see more people becoming capable of building and launching features independently. But we’ll also need people who are equally skilled at recognizing when a feature has outlived its usefulness.</p><p>That judgment will only become valuable if organizations are willing to recognize and reward it. Otherwise, faster development could lead to larger, more complicated products.</p><p>There’s also another consideration that deserves more attention: the difference between the cost of building a feature and the cost of maintaining it. Launching a feature is only the beginning. Once it’s part of a product, it may require ongoing updates, support, testing, documentation, and moments whenever other parts of the system change.</p><p>Some features may eventually require more effort to maintain than the value they continue to provide. For that reason, product managers should look beyond how quickly or cheaply something can be built. They also need to consider what it will take to support that feature over time and whether it continues to justify those resources.</p><p>A feature that was worth building yesterday may not necessarily be worth maintaining tomorrow. Recognizing that difference, and being prepared to act on it, could become one of the more important responsibilities of product management as AI changes the pace of development.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HQwLEweM05AoOG4xDsk8xA.png" /><figcaption>Tea ceremony.</figcaption></figure><h3>Attention becomes the thing you run out of</h3><blockquote>There’s something remarkable about the quiet discipline of a Japanese tea ceremony. As pictured above, preparing a bowl of tea becomes an exercise in patience, concentration, and giving one task your full attention. Nothing is rushed, and every movement is deliberate. Product teams may find this increasingly difficult as AI allows them to manage more ideas, experiments, and development activities at the same time.</blockquote><blockquote>While AI can work on several tasks simultaneously, our ability to concentrate, evaluate results, and make thoughtful decisions remains limited. The question is no longer just how much work we can produce, but how much of it we can properly understand, manage, and give the attention it deserves.</blockquote><p>For years, the need to prioritize product ideas also helped keep workloads manageable. Development teams could only build so much at a time, which naturally limited how many projects could move forward simultaneously. Nobody necessarily designed the process to protect people’s time and attention, but that’s partly what it did. As AI removes some of those development limitations, we may also lose the natural boundaries that once helped keep workloads under control.</p><p>Imagine a product manager working on several ideas at once, each supported by AI tools such as Claude Code or Codex. One feature might be ready for testing, another could need additional instructions, while a third is waiting for someone to review the results. Each requires careful attention, even though the AI systems are doing much of the actual development work.</p><p>And because these tools can continue producing work without needing the same breaks we do, there’s always another task to review, another question to answer, or another improvement to consider.</p><p>The difficulty is that people cannot divide their attention across multiple complex tasks as easily as AI systems can process them. Moving constantly between different projects requires mental effort, and the more frequently we switch between them, the harder it becomes to give each one the attention it deserves. This isn’t just a question of managing stress or maintaining a healthy work-life balance, although those concerns are important. It also affects the quality of product decisions.</p><p>AI may be able to generate ideas, build prototypes, and produce results at a much faster pace, but people still need to examine that work, understand its implications, and decide what should happen next.</p><p>If product managers become overwhelmed by the amount of work coming their way, their ability to make thoughtful decisions can suffer. The technology may continue performing well, but the organization may struggle to keep up with what it’s producing.</p><p>This introduces another form of prioritization that I believe will become increasingly important:<em> knowing when to stop work that is already underway. </em>Sometimes that means putting an experiment on hold, closing an AI-assisted development task, or deciding not to pursue another idea because there isn’t enough time or attention to evaluate it properly.</p><p>These decisions may not attract much recognition because nothing has been launched or visibly removed from the product. But they can be just as valuable as deciding which existing features should stay or go.</p><p>Ultimately, the ability to produce more work doesn’t automatically mean we should take on more work. As AI increases what product teams can accomplish, knowing where to focus our limited attention may become one of the most important decisions we make.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oHJ5rOxh0sVEj7XI7S6JPA.png" /><figcaption>Miegakure garden.</figcaption></figure><h3>What happens to roadmaps</h3><blockquote>A winding path through a Japanese garden doesn’t always reveal where it leads. As pictured above, each turn offers a new view, while the path itself provides a sense of direction. This reflects an important shift in how product teams may approach planning.</blockquote><blockquote>Rather than committing too early to a fixed list of features, teams can focus on the results they want to achieve and remain flexible about how they get there. As AI makes it easier to explore and test different ideas, product managers may have more freedom to adjust their approach along the way, without losing sight of the outcome they’re working toward.</blockquote><p>If AI makes it easier and faster to build software, what does that mean for the way product teams plan their future work? For nearly fifteen years, I’ve wanted to work with outcome-based roadmaps, where the focus is on what we’re trying to achieve for customers and the business, rather than committing to a fixed list of features. But in practice, I never quite managed to work that way.</p><p>The problem wasn’t that people disagreed with the idea of focusing on outcomes. It was that building software took time, and engineering resources were limited. If a team wanted to accomplish something, it usually had to identify the specific features it planned to build well in advance, simply to secure the development time needed to deliver them.</p><p>As a result, even when roadmaps were presented around business goals and customer outcomes, they often ended up being lists of features with expected delivery dates. The outcomes were there, but the actual planning still revolved around what needed to be built. AI could finally give product teams an opportunity to change that approach.</p><blockquote>If building and testing features becomes faster and less expensive, teams may no longer need to commit to every solution months in advance. Instead, they can establish a clear outcome, explore different ways of achieving it, and use customer feedback and performance data to guide their decisions.</blockquote><p>For example, rather than committing to building several specific features to improve customer engagement, a team could focus on the goal of helping customers return to the product more frequently. It could then experiment with different ideas, measure what works, improve promising solutions, and abandon those that don’t produce meaningful results.</p><p>This would allow roadmaps to become more flexible while keeping teams accountable for what they’re trying to accomplish. Of course, that doesn’t mean traditional roadmaps will disappear.</p><p>Some projects will still require significant engineering effort, coordination across departments, substantial investment, and careful long-term planning. AI may help with parts of that work, but it won’t eliminate every dependency or make every complex project easy to deliver.</p><p>I expect the future will involve a combination of both approaches. Organizations will continue to make firm commitments for work that requires considerable planning and resources, while allowing greater flexibility in areas where teams can experiment and adjust their solutions along the way.</p><p>The important shift is that product teams may finally have more freedom to commit to the results they want to achieve, without having to decide every feature they’ll build before the work even begins.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MPCH1X0xgg1GQkeK-WCKhA.png" /><figcaption>Kintsugi bowl.</figcaption></figure><h3>Why I’m up for this trade</h3><blockquote>In Japan, the art of Kintsugi involves repairing broken pottery with lacquer and gold, turning visible cracks into part of the object’s beauty rather than hiding them. The restored bowl pictured above reflects the idea that something can change, even experience damage, and still retain its value.</blockquote><blockquote>The same may be true for product management as AI reshapes how we work. Some of the methods we’ve grown accustomed to may change, but that doesn’t mean the experience and judgment we’ve developed along the way become less valuable. Instead, they may take on new meaning as we adapt to different ways of building products and making decisions about their future.</blockquote><p>The pace of change will differ from one company to another, but I believe the overall direction is becoming clear. AI is making it possible to develop software faster, and organizations that have spent years trying to overcome development limitations are unlikely to ignore that opportunity.</p><p>What interests me most, however, isn’t simply the ability to build more. It’s what that freedom could mean for the ideas we’ve traditionally had to leave behind. For years, promising ideas were often set aside because there wasn’t enough time, money, or engineering capacity to explore them. As those limitations begin to ease, product teams may have more opportunities to experiment, learn directly from customers, and discover valuable solutions that might otherwise never have been developed.</p><p>Product managers may also have greater opportunities to demonstrate the quality of their judgment, not simply through what they build, but through the decisions they make, the results they achieve, and their willingness to change direction when something isn’t working.</p><p>Of course, these changes will affect how people work, and not everyone will find the transition easy. Responsibilities will evolve, familiar processes may disappear, and some of the skills we’ve spent years developing may need to be applied differently. But not every change means losing what made something valuable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tvIqzDQrLUTPUjsVogYrfw.png" /><figcaption>Kintsugi Master in a Serene Tatami Room</figcaption></figure><h3>The work changes. The purpose remains.</h3><blockquote>The Japanese artisan pictured above demonstrates Kintsugi, the careful restoration of pottery using gold to highlight rather than hide its imperfections. As AI transforms product management, the tools and methods we use may evolve, but the experience, judgment, and creativity we’ve developed remain essential. Our craft may change, but the human purpose behind it remains.</blockquote><p>It’s another reason why I don’t believe the fundamental purpose of product management is changing. Our responsibility has always been to understand customer needs, make thoughtful decisions, and help create products that deliver real value. AI may change the tools we use and the speed at which we work, but those responsibilities remain just as important. If anything, having greater freedom to build may place even more responsibility on our ability to decide wisely.</p><blockquote>We may spend less time worrying about whether something can be built and more time considering whether it should be built, how well it serves customers, and whether it deserves to remain part of the product.</blockquote><p>For those of us who value the purpose of product management more than the particular methods we’ve grown accustomed to, there’s something encouraging about that. We have an opportunity to spend more of our time understanding problems, exploring possibilities, and making decisions that genuinely improve the products people use.</p><p>And for those who find that the work they once enjoyed is changing, there may be opportunities to discover different roles where their experience and abilities can continue to make a meaningful contribution.<em> You could call that a freedom.</em></p><p>These reflections come from what we’re beginning to experience at Livesport, and I recognize that other organizations may be seeing things differently. None of us can say with certainty how all of this will unfold, but it leaves me wondering: <em>if AI gives us the ability to build almost anything, how will we become better at deciding what is truly worth building?</em></p><p>And as these tools become more capable, will we use that freedom to create better products, or just end up creating more of them? I believe these are questions worth exploring together, because the future of product management may depend less on what technology allows us to build and more on the judgment we bring to deciding what really matters.</p><p><em>Many thanks to </em><a href="https://medium.com/@tremisskeete"><em>Tremis Skeete</em></a><em>, Executive Editor of Product Coalition, for his invaluable contributions to this story. A special shout-out to founder </em><a href="https://jaystansell.medium.com/"><em>Jay Stansell</em></a><em> for creating an environment that enhances product management education.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a12ca5636c16" width="1" height="1" alt=""><hr><p><a href="https://medium.productcoalition.com/you-can-ship-anything-now-can-you-kill-anything-a12ca5636c16">When Building Gets Easier, Decisions Get Harder</a> was originally published in <a href="https://medium.productcoalition.com/">Product Coalition</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lessons From 13 Claude Code Projects]]></title>
            <link>https://medium.productcoalition.com/13-claude-code-projects-that-changed-my-product-manager-role-over-the-last-6-months-7057b9045d51?source=rss-a94f151b844d------2</link>
            <guid isPermaLink="false">https://medium.com/p/7057b9045d51</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[projects]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <dc:creator><![CDATA[Ondrej Machart]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 21:21:54 GMT</pubDate>
            <atom:updated>2026-09-30T00:46:35.831Z</atom:updated>
            <content:encoded><![CDATA[<h4>And how they changed my mindset and role as a Product Manager.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TK6Bn1WgKma5hSD08-wRLw.png" /></figure><p>It’s been just a few days since the one-year anniversary of the term “vibe coding” — coined by <a href="https://x.com/karpathy/status/1886192184808149383?lang=en">Andrej Karpathy</a> on X. That surprised me. It feels like it’s been around for at least two or three years. The speed of change lately is fascinating. I was a Head of Product a year ago and I still am — but the role has changed so much I can barely recognize it anymore.</p><p>Where it used to be about influencing future and imminent changes through complex collaboration and orchestration of various people across domains, it now revolves a lot around influencing through building and creating, mostly with the aim to speed up change thanks to vivid prototypes or even delivery of working products.</p><p>I decided to look back and show you 13 Claude Code projects that made me a different PM and that might either spark imagination on your end or a discussion about your own transition experience. Both work.</p><p>👋🏻 My name’s <a href="https://www.linkedin.com/in/ondrejmachart/">Ondrej</a>. I’m a Product guy. I used to be a UX designer. I’m not an engineer. This is my story of learning to build with AI.</p><h3><strong>Overview of the Claude Code Projects</strong></h3><p>Spanning websites, web apps, iOS apps, custom MCP, Machine-learning prediction model, data analysis, presentations and even a print project.</p><ol><li><strong>🧭 Little Explorer</strong> — Native iOS app for parents to find playgrounds and family-friendly places nearby</li><li><strong>⚙️ Little Explorer Ecosystem</strong> — Web admin for remote app configuration and macOS tool for testing map data across continents</li><li><strong>👋🏻 Ondrejmachart.com</strong> — Personal website built in minutes, styled over hours</li><li><strong>🗺️ Product Portfolio Coach</strong> — Web app that mapped 30+ Livesport products and helped leadership make a tough consolidation call</li><li><strong>📱 Flashscore iOS Prototype</strong>— A vibe-coded clone of Flashscore iOS app so colleagues can prototype on top of it</li><li><strong>🤖 ProductHub</strong>— Custom MCP I built to learn how MCPs work — and then stopped using</li><li><strong>🌳 Family Tree</strong> — Printed family tree that goes back to the 17th century, made for my dad for Christmas</li><li><strong>⚽️ Football Predictions</strong> — An ML model that tried to predict football goals — and taught me more than the predictions did</li><li><strong>💡 Vibe-Coding Workshops</strong> — Helping colleagues at Livesport start building with AI</li><li><strong>🍼 Tinystakeholders.com</strong> — Lenny Rachitsky’s podcast transcripts turned into parenting hacks for tired PMs</li><li><strong>🎨 Infographics Tool Prototype</strong> — Configurable sports infographics for editorial teams — scary close to production</li><li><strong>📊 Data Analysis &amp; Presentations</strong> — Using Claude to analyze data, create web presentations and Confluence pages</li><li><strong>✅ Personal Task Tracker</strong>— Because no tool on the market serves my needs perfectly</li></ol><p>Some are still running in production. Some helped our C-level make a major consolidation decision. One is just a family tree I made for my dad.<br>I’m not sharing this to sell you a course or convince you AI will save your job. I’m sharing because this transition has been equal parts exciting and terrifying, and I think we need more honest conversations about what it actually takes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sDith8vjaUimM9T0IHFxRQ.jpeg" /><figcaption>“There’s a new kind of coding I call “vibe-coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” — Andrej Karpathy</figcaption></figure><h3><strong>🏄 Mindset Shifts That Can Get You Further</strong></h3><p>Before we get into the projects, here are the mindset shifts that made the biggest difference along the way. I wish someone told me these before I started. The projects below will show you what they look like in practice.</p><p>🔧 <strong>It’s not magic, it’s work.</strong> AI gets you far but it’s not free, not effortless, not without risk. It’s still a lot of work, just different than coding.</p><p><strong>🥚 Start by solving your own problem.</strong> Learn the basics on your own case before trying to solve someone else’s problems. You’ll get immediate feedback from your most devoted user and fan.</p><p><strong>🪦 Your project can die and it may still be a win.</strong> Some of my projects served their purpose and didn’t need to live on. The outcome isn’t always the tool. It can be learning something or influencing a decision.</p><p>🏔️ <strong>Respect the challenge you’re tackling. </strong>Vibe-coding does not equal competence. The gap between prototype and production is real and humbling.</p><p>👀 <strong>Know your blind spots.</strong> A UX designer might vibe-code fantastic frontend experience with a backend that won’t scale or will be leaky. An Engineer might vibe-code a functional product that can scale but won’t because it’s unusable. Both can fail because the decide to tackle a problem nobody wants to pay for having solved. Know your blind spot and think about your options to remove it.</p><p><strong>🛋️ Don’t protect your routine from new things. </strong>AI comes with something new every week or two. Be open to it. Don’t jump on anything and everything immediately, but also don’t close yourself off from new things entirely when the comfort zone gets established.</p><p>🤝 <strong>Let others take over and surpass you.</strong> You don’t always have to be the solo builder, even if that feels more natural to you. I strongly believe AI is not here to make us more individualistic, just more lean in terms of how many people are needed to build something. Partner up with someone who also loves building with AI and who has the skills you don’t. Choose someone you enjoy working with.</p><p>🎲 <strong>Embrace AI’s non-determinism.</strong> Stop fighting the randomness — sometimes it’s exactly the right tool for creative work.</p><p>And most importantly, <strong>have fun with it</strong>. Taking it too seriously too early may just kill the vibes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sFY34EVGTGG1cD56OevwxA.jpeg" /><figcaption><strong>Little Explorer</strong> — Native iOS app for parents to find playgrounds and family-friendly places nearby</figcaption></figure><h3><strong>1. Little Explorer — iOS App Published to the App Store</strong></h3><p>This was in September 2025. So just six months ago. Before this project, I used AI mostly to polish copy, improve emails, generate images, research markets. Optimization of outputs rather than creating complex things.</p><p>I had a mind full of ideas for my own little pet projects. So I obviously tried Replit, Lovable, and all the other vibe-coding tools and created a few prototypes. But none at the time offered a way to build an actual native mobile app (not true anymore).</p><p>Then a colleague of mine suggested using Claude Code in the terminal, letting it basically work with an Xcode project and build the app. I fell in love with the idea and for two months I learned the ropes over evenings and nights. Motivated not by getting rich overnight but by building something people might actually use. Completely solo.</p><blockquote>Nothing beats the stress of changing something that can break on people’s phones when it’s your name under it.</blockquote><p>In November 2025, I finally published the <a href="https://apps.apple.com/cz/app/little-explorer/id6749822199">app in the App Store</a>. With a subscription built in, with several third-party API integrations, and with about a hundred users monthly — parents who need to find the nearest playground, toilet, or other family-friendly places nearby.</p><p>This was a kick-start for me because it wasn’t just a prototype. <strong>I had to deal with securing the app, lots of debugging, API provider implementation, App Store submission, legal document creation, etc. </strong>It was so much fun. And it just felt good to see the app in the store. Because in a world of a PM who can spend months pushing for changes, this felt like each day had an instant reward moment, or several of them. Something a creative mind craves.</p><blockquote>👉 If there’s a takeaway for you here, I think it’s that AI can get you as far as publishing in the App Store — but it’s not for free, it’s not without effort, it’s not without a lot of learning, it’s not without the risk of failing, and it’s not worry-free in terms of the quality of the code if you’re not technical enough to keep an eye on the architecture. It takes a lot of iterations of how you approach the coding.</blockquote><p>PS. Here’s an <a href="https://medium.com/@ondrej.machart/from-zero-to-app-store-building-a-complex-native-ios-product-with-ai-42b45626f16e">article about building Little Explorer</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZNAlUhnx2ulJ21Jl9Cfrfg.jpeg" /><figcaption><strong>Little Explorer Ecosystem</strong> — Web admin for remote app configuration and macOS tool for testing map data across continents</figcaption></figure><h3><strong>2. Little Explorer Ecosystem — Remote Config and macOS Admin</strong></h3><p>As part of the Little Explorer project, I soon realized that optimizing the API configuration for finding the nearest playground, toilet, zoo, or picnic area requires unreasonably frequent app updates. <strong>So I decided to take this part of the app’s business logic and move it to a web application.</strong> There, I can remotely configure each of the nine categories in the app — which API provider to use, how to search, what radius to apply. All of this can be changed and the apps in people’s phones will pick it up on the next launch — not after updating the app.</p><p>This was tricky because it’s not just about creating a nice interface. <strong>This is a critical dependency for an app that is on the App Store and in people’s devices.</strong> I had to learn a lot about testing to make sure it’s bulletproof. I still wake up at night with ideas where it might be weak. Nothing beats the stress of changing something that can break on people’s phones when it’s your name under it. That stress is healthy — it makes me double down on testing and learning.</p><blockquote>That stress is healthy — it makes me double down on testing and learning.</blockquote><p>I also built a macOS admin app — partly because I was curious if I could, partly because I genuinely needed it. I wanted to test the API providers across locations — does MapKit work for toilets only in the US or in Europe too? Can Mapy.com find good playgrounds not just in Prague but also in Jakarta? The macOS app lets me configure the API, see results on an interactive map, and move between cities, villages, continents. It also lets me see all the places found by real users and check if the playgrounds it found are actual playgrounds or something completely different — like a junction of three highways in Singapore (happened) or a WWII monument (also happened). Because it’s a macOS app, it can test the MapKit API natively, which behaves completely differently on web.</p><blockquote>👉 The takeaway from this whole ecosystem is that AI can get you surprisingly far beyond the prototype — into production dependencies, remote configuration, cross-platform testing. But the further you go, the more you need to respect what you’re building and who depends on it.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EN7d2IKxf1YdrJrFLEQwSg.jpeg" /><figcaption><strong>Ondrejmachart.com</strong> — Personal website built in minutes, styled over hours</figcaption></figure><h3><strong>3. Ondrejmachart.com — Personal Website</strong></h3><p>Well, this one is easy. There’s probably no better and safer use of Claude Code or any coding AI model than building <a href="https://www.ondrejmachart.com/">your own little static website</a>. <strong>Hosted for free on Vercel, built within minutes.</strong> I needed a place to lead people from the App Store — about me, contact, app landing page, legal documents. Simple needs.</p><p><strong>The fun realization was that while the vibe coding took minutes, it took hours to iterate the little details like font, style, colors, etc. </strong>Because this time I had absolutely no vision, just a need. But once the website appeared in front of me, I suddenly started to have opinions about how it looks — unstructured, chaotic, changing, unclear.</p><p>Frankly, during vibe coding of most projects I don’t mind this state of mind because the UI appears rough at first and I slowly but surely mold it through using the prototype and iterating it. But when it comes to a static personal website, it feels more like building a digital mirror — and that’s somehow harder to wing (or, well, vibe).</p><blockquote>👉 So the surprising takeaway is that if you ever vibe code your personal website, maybe don’t start with nothing but the purpose and content (that should always come first, period) — spend a bit more time prepping about the style. You will want to feel good about it. And vibing your way to a style is exhausting.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*df2ZAS1PMbuG9POggYN_cA.jpeg" /><figcaption>Product Portfolio Mapper — Product logos are masked to keep the company internal details protected.</figcaption></figure><h3><strong>4. Product Portfolio Coach — Web App for Livesport</strong></h3><p>This was the first “serious” tool I vibe coded at work. Our company has over 30 products in its portfolio. Some similar, some specialized, some larger, some smaller. <strong>What we always lacked was a space where you could look at all the products through various lenses</strong>: finding the most similar products ripe for consolidation, making decisions about differentiation of new products, etc. So I decided to make a tool that would give us this level of transparency and clarity, to help us make better decisions through seeing the forest, not just the trees.</p><p>It was fairly simple in terms of data — I didn’t connect it to any internal API, I just created an admin section where I could manage the product information. Where it got interesting was the frontend and interaction UX. I wanted to create a canvas where one could freely move the products on two selected axes, e.g. product size and primary audience. Achieving a FigJam-like experience wasn’t easy, but I got pretty close to it with optimistic change handling.</p><blockquote>What made my day was that the tool eventually helped our C-level management make a tough decision about major brand consolidation</blockquote><p>I also wanted to create a “similarity map” where one could see clusters of products that share a lot of traits — like the sports they focus on, the depth of sports data, the complexity or simplicity of the data visualizations, etc. It was both fun to create a complex UI and to uncover the insights about the portfolio. I also added an AI feature that allows us to pick any product(s) and let AI analyze them through all the data we gathered in the tool.</p><p>What made my day was that the tool eventually helped our C-level management make a tough decision about major brand consolidation that aims to help the company work more effectively towards its ambitious goals. <strong>With fewer products, this tool has a dim future — but that’s OK. It served its purpose, and not needing it is actually a great outcome.</strong></p><blockquote>👉 I would say this is one of the takeaways — if you see a problem worth tackling and dare to try vibe-coding your way out of it, give it a shot. You don’t know how far it will get you, but you will definitely spend quality time analyzing the problem, which is healthy — no matter whether the tool becomes the solution or just a vehicle to get you to it.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xBX_yEUOp2j63vW_NVwNDQ@2x.png" /><figcaption>Flashscore iOS shared repo — A vibe-coded clone of Flashscore iOS so colleagues can prototype on top of it</figcaption></figure><h3><strong>5. Flashscore iOS Prototype — A Shared Base for Prototyping</strong></h3><p>The next challenge I wanted to tackle was massive: to vibe-code a clone of the Flashscore iOS app so that me and my colleagues all have a great starting point when prototyping new features and ideas. <strong>The problem I wanted to solve was that whenever people tried to prototype using vibe-coding tools like Replit, Lovable, or others, they had to first spend a fair portion of time and effort cloning the UI. </strong>They provided AI with screenshots, Figma files, exported styles — but it got them mostly nowhere. That’s fine for quick exploration, but when you want to test with users or show stakeholders, you need to get closer to the real product.</p><p>This turned out to be a perfect challenge for my awesome colleague <a href="https://www.linkedin.com/in/petrajavorkova/">Petra</a>, who is our DesignOps team lead — she started to create one screen after another, providing the team with a shared GitHub repo anyone could pull to start working on their prototype. <strong>Because Petra is more technical than I am, she completely took over and marched forward.</strong> We’ve already prototyped a couple of ideas like a top matches widget, and it works (almost) like magic thanks to the Figma MCP — you copy a prompt, it takes UI from Figma, and you have the prototype ready, sometimes on the first try. But it works great also when we don’t have any Figma design yet and just want to explore ideas.</p><blockquote>👉 The takeaway here is that sometimes vibe-coding can be a tiny quick project, but sometimes it can become a multi-month challenge to enable others to prototype quickly — and that it’s great to have someone who’s more suited for that challenge and is up for it.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*19VU56cEcpdS2F-uTWJ4oA@2x.png" /><figcaption>Custom MCP I built to learn how MCPs work — and then stopped using</figcaption></figure><h3><strong>6. ProductHub — Custom MCP with Kanban Board</strong></h3><p>At this point I felt a bit weird. I had gotten quite comfortable with my setup: Claude Code in the terminal, directly building on my file system. I alternated types of projects and platforms, but it started to feel like a routine. Then MCPs became a thing. I heard about them on podcasts, read about them in newsletters. And somehow I couldn’t wrap my head around them. They felt new and kind of threatening to my routine.</p><blockquote>If AI is coming for our jobs, we should at least also have some fun with it while we can.</blockquote><p>When I finally decided I was open to learning about MCPs, I asked Claude what’s the best way to do so. <strong>Claude suggested building my own MCP. That sounded pretty cool — I love learning by doing.</strong> I decided to build a hub for vibe-coding projects — a Jira-like kanban board where I could create tickets, and a custom MCP that would enable Claude to connect to the hub, to the respective project, and to the tickets.</p><blockquote>I decided to build an MCP for vibe-coding projects — a Jira-like kanban board.</blockquote><p>And it worked! I could tell Claude to work on five of the tickets in the backlog, moving them to “for review” when done, committing to git separately. <strong>I went to lunch and when I returned, the tickets were all waiting for <em>me</em>. </strong>So I added another tool — to save important context about the project so that next time I start a session, Claude can pull all the important info via the MCP. That worked too, albeit it uncovered a new layer of complexity. Maintaining the context and avoiding confusion wasn’t easy.</p><blockquote>Eventually I realized I actually prefer not to use my own tool. <strong>Creating tickets felt unnatural to the vibe-coding routine.</strong></blockquote><p>Eventually I realized I actually prefer not to use my own tool. <strong>Creating tickets felt unnatural to the vibe-coding routine. </strong>It added steps to what benefits from a lack of process when exploring ideas. I decided the purpose of this project was served — I learned how MCPs work. A few months later I discovered Serena MCP — context management that works without my inputs and extra steps. Plus, most task management tools published their own MCPs, so if I ever decide I need a bit more organization to my flow, the solution is out there.</p><blockquote>👉 The takeaway here is simple: there’s something new in the realm of AI every week or two. The best way to learn the new tricks is not to read about them but to use them.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x6sM_QIJExD3XL_7vGtX0g.jpeg" /><figcaption>Printed Family Tree Going Back to the 17th Century</figcaption></figure><h3><strong>7. Family Tree — Printed Family Tree Going Back to the 17th Century</strong></h3><p>My father had a firm belief that our family comes from a French soldier from Napoleon’s army who fled and knocked up a Bohemian girl. I’m not sure how he arrived at this, but he was dedicated enough to prove it that he hired a genealogist to track our true roots. She made it to the 17th century. No French soldiers, just poor tailors of questionable morale. This was a bit of a flop for him.</p><p><strong>To cheer him up, I decided to turn the genealogist’s notes into a proper family tree </strong>— with a family coat of arms, a map of places where our ancestors lived, their stories woven into the historical context, a tree design, and AI-generated portraits of our ancestors based on photos of our family members. All generated with AI. Images made by ChatGPT, text by Claude. I let Claude read the notes, do online research on the history of the villages where my great-great-great-grandpas lived, and then I put it all together in Figma (which was surprisingly OK for print).</p><p>I had it printed on A0. My father loved it so much he decided to print it on an even larger format and started giving it to his cousins and distant relatives — even those who aren’t captured on the family tree. It truly made his day to see little details related to his professions and passions, and his dad’s. And for me it was a nice pet project after all the work stuff.</p><blockquote>👉 The takeaway wasn’t the print. What I realized is that I often get frustrated when building with AI when it doesn’t behave somewhat logically. But AI is in its nature non-deterministic. It is made to be creative, to act somewhat unexpectedly. This was a perfect fit for its capabilities.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1fDhaj7bg3zU7UeJxyUq6A.jpeg" /><figcaption>Football Predictions — ML Model for Major European Leagues</figcaption></figure><h3><strong>8. Football Predictions — ML Model for Major European Leagues</strong></h3><p>Our company has always been fanatic about sports data. Then we got obsessed with insights. And lately, we’re falling in love with the idea of predictions. Not subjective ones, but data-powered. Predicting goals is like the ultimate pinnacle of drawing insights from the wealth of stats and data.</p><p>That got me thinking about two things. First, how would predictions look and feel in the product? And second: <strong>hell, can I build a machine-learning model predicting goals?</strong> Never mind that I wouldn’t be the first or last one to attempt it. The fun element of giving it a shot was just too strong to resist. So I had Claude research the methods of predicting football, then I had a discussion with it to help me understand it a bit more. And then, as usual, we got to work.</p><blockquote>I did not make a model that beats bookmakers. But I explored tons of ways to work with the data in the UI that helped me as a PM to have fruitful sessions with the UX designer, AB tester, and engineers.</blockquote><p>After a couple of iterations, I had some interesting predictions for the Premier League to test against reality. And I also learned about feature engineering, time decay, MAE, LightGBM, hyperparameter tuning, walk-forward validation, and XGBoost. Things that probably won’t make me a fun person to chat with at a party (unless it’s a nerd one) but that <strong>excited me and made me appreciate the complexity of the ML domain.</strong></p><p>Did the predictions work? Well, let’s just say that football is beautiful because even bookmakers who have the most advanced ways of predicting match outcomes are only slightly better than a flip of a coin. I did not make a model that beats them. But I explored tons of ways to work with the data in the UI that helped me as a PM to have fruitful sessions with the UX designer, AB tester, and engineers. <strong>That’s more than <em>I</em> predicted.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xX2IbY52gzGsxN-KbdzmFA.jpeg" /><figcaption>Vibe-Coding Workshops — For Colleagues at Livesport</figcaption></figure><h3><strong>9. Vibe-Coding Workshops — For Colleagues at Livesport</strong></h3><p>This AI era is pretty stressful for PMs (and not just for them), to be honest. We have a lot on our plates any day of the week. A lot of operational work that won’t wait, a lot of conversations that should have happened yesterday, a lot of stuff to ship. <strong>To allow yourself to slow down or even stop for a while to catch up with everything that’s been happening with AI feels impossible most of the time. </strong>To adopt the tools and new ways of working may feel out of your comfort zone, slowing you down at first, risking the quality of output people expect or depend on.</p><p>I found myself learning a lot only thanks to spending time vibe-coding and playing with AI in late hours, when my wife and kiddo sleep. It’s not healthy and sustainable. But somehow it felt impossible not to do it this way. It felt stressful to fall behind. But perhaps more importantly, it also felt the most exciting thing since my discovery of Lego as a kid. Those bricks gave my imagination an endless space for creativity (well, limited by my parents’ Lego budget). AI feels the same way.</p><blockquote>There was never a lack of ideas to build, just the barrier to start building.</blockquote><p>Thanks to this — and thanks to a more healthy attitude towards sleep time on my colleagues’ part — I got a bit ahead in the use of AI. <strong>So I soon began organizing workshops for those interested in how I build with Claude Code.</strong></p><p>Most colleagues who joined used AI, just not for building stuff. Turned out it was mostly just because they didn’t know how to set it all up — Claude Code in the terminal, the packages, etc. <strong>There was never a lack of ideas to build, just the barrier to start building.</strong> Most colleagues left with a prototype of their idea. Some came to me later with something real:</p><ul><li>A <a href="https://apps.apple.com/cz/app/brainbatch/id6757217874">morning quiz game app</a> in the App Store by <a href="https://apps.apple.com/cz/app/brainbatch/id6757217874">Jirka</a>.</li><li>A personal planner app for a big Japan trip.</li><li>A website with <a href="https://www.pojdsemkamjdes.cz/">gourmet restaurant tips</a> by <a href="https://www.linkedin.com/in/martin-n-a3a59993/">Martin</a>.</li><li>An internal tool for UX team capacity allocation management.</li><li>An internal tool for design system implementation monitoring by <a href="https://readyforops.design/">Petra</a>.</li><li>A <a href="https://www.horakvodotopoplyn.cz/">website</a> for dad’s plumbing business by <a href="https://www.linkedin.com/in/kristynasloukova/">Kristyna</a>.</li><li>A Pacman game where Pacman is a cat trying to make its way through rooms where roomba, angry kid and other enemies get in the way</li></ul><p>So many great ideas. It felt awesome to see people giving shape to the solutions they craved.</p><blockquote>👉 This eventually sparked a change in my role. Where I used to be a bit drowned in operational delivery details, I see my role becoming more focused on tackling complex challenges through rapid prototyping to shine a light on the ways forward. And what makes me truly happy is to see my colleagues picking up the tricks, oftentimes soon surpassing my own skills in their own ways.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FzvI6dFhHLRv1zqIayIS-Q.jpeg" /><figcaption>TinyStakeholders.com — Lenny Rachitsky’s podcast transcripts turned into parenting hacks for tired PMs</figcaption></figure><h3><strong>10. Tinystakeholders.com — Product Wisdom Meets Parenting Hacks</strong></h3><p>Early in 2026, my favorite podcast author <a href="https://www.lennysnewsletter.com/">Lenny Rachitsky</a> made all his 300 episode transcripts available for the public to play with. What a nice move from him. Immediately, people started to share amazing projects, mostly revolving around using the sheer power of insights Lenny’s guests have shared throughout the years to answer product people’s burning questions.</p><p><strong>My burning questions lately have been “why is my 2-year-old son running around naked refusing to dress up,” “when will he stop licking things,” and “what’s this weird thing on the wall.” </strong>I thought it might be fun to try to use Lenny’s transcripts to draw parenting wisdom (or at least some cheap tricks) rooted in PM concepts, frameworks, and tactics. To my surprise, it worked pretty well. It felt weirdly scientific, so I let Claude learn the humor of some famous standup comedians and rewrite the parenting tips to make the tired PM parent laugh-cry into their glass of wine at 2 AM.</p><p>I sent <a href="https://www.tinystakeholders.com/">TinyStakeholders.com</a> to Lenny, he posted it on X and LinkedIn. I wrote a <a href="https://medium.com/@ondrej.machart/from-vibe-coding-to-tinystakeholders-an-effective-guide-for-the-non-technical-solo-builder-8125c0fac25b">Medium article</a>, he loved that too. That genuinely made my day. Over 5,000 people read the article and saw the website, some sent me nice messages about how it motivated them to build something too.</p><p>I have zero dollars from this project. It has nothing to do with my job. Lenny did not invite me to his podcast. Most of the parenting tips did not help to get my son to enjoy dressing up. But I had a hell of a lot of fun learning how to work with a large dataset (300 episodes, each roughly 90 minutes — that’s a lot of words!) and adding that baby with a toilet paper roll to the homepage.</p><blockquote>👉 The takeaway is that if AI is coming for our jobs, we should at least also have some fun with it while we can.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6D4nBzV0wS-9H9sS85R3fw.jpeg" /><figcaption>Livesport Infographics — Configurable sports infographics for editorial teams, scary close to production</figcaption></figure><h3><strong>11. Infographics Tool Prototype — Scary Close to Becoming an Internal Tool</strong></h3><p>Things can get awkward with vibe-coding. When done wrong, it fails spectacularly. By done wrong I mean treating vibe-coding as magic — building things you don’t understand, not learning how they work, and assuming that a functioning prototype makes you a competent builder. This is the case with most projects people build — not because they’re stupid or lazy, but because <strong>curiosity alone doesn’t always carry you through the extra mile</strong>. And even if you try hard, without technical knowledge there will always be the gap and the risk that comes with it.</p><blockquote>Things can get awkward with vibe-coding. When done wrong, it fails spectacularly.</blockquote><p>This is a long prologue to this project. <strong>Because it’s the first project that seems to be evolving from a vibe-coded prototype into a production internal tool.</strong> It raised questions, eyebrows. It made me feel like an idiot when pitching the idea to skilled developers and TPOs. It’s too soon to say if we made the right call because we’re still building it, not sure if it’s the way forward. But at least I feel better now because a great colleague, a senior engineer, joined forces and took the leap of faith with me to orchestrate the architecture and data plumbing while me and my PM colleague vibe-code the frontend.</p><p>The tool offers configurable infographics for our editorial teams to enrich their sports articles. It auto-translates, exports images, works in three different aspect ratios, and supports multi-brand usage. It’s scary. That’s good. <strong>It’s not a production tool yet and we’re not sure if it will be. But we are exploring the path. And that’s perhaps more important than whether this will become the official tool or not.</strong></p><blockquote>👉 Vibe-coding needs more respect — not more respect for vibe-coding from people, but more respect from vibe-coders towards the challenges they dare to tackle. Take that as a takeaway.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oZEvyHOqDGnkRmghXSAK_Q.jpeg" /><figcaption>Data Presentations — Using Claude to analyze data, create web presentations and Confluence pages (numbers are masked to protect company’s data)</figcaption></figure><h3><strong>12. Data Presentations — Web Presentations and Confluence Pages</strong></h3><p>The first data analysis Claude Code did for me had great output content-wise, but design-wise it was just an ugly standard terminal message. I pasted it into the slides, but I knew there’s more Claude Code could do with the data and insights. <strong>So next time I asked Claude to take the data and insights and turn them into a web presentation</strong> — literally coded slides, with keyboard controls, no scrolling, Livesport brand applied.</p><blockquote>I asked Claude if we could turn the web presentation into a Confluence page. It suggested using a community MCP. It worked just fine</blockquote><p>It was great. We had a decision to make about sunsetting the support for Huawei devices and we managed the internal communication smoothly thanks to the presentation Claude created. <strong>A chart that showed less than 1 percent share of our audience made everyone relaxed. </strong>Except my boss Tomas, who prefers to have these things either in Slides or on Confluence. So I asked Claude if we could turn the web presentation into a Confluence page. It suggested using a community MCP. It worked just fine — the decision is now documented on our Confluence space with all the charts exported and insights captured.</p><p>When my documentation-obsessed colleague saw this, he almost fell off his chair. He loves having things documented but hates documenting them. He’s a free man now.</p><blockquote>👉 Take that away — explore the MCPs of your favorite tools. Chances are Claude can do so much more for you than you think.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iZA9pXg1CAexU9R-CXZJDw.jpeg" /><figcaption>Personal Task Tracker — Because No Tool Serves My Needs Perfectly</figcaption></figure><h3><strong>13. Personal Task Tracker — Because No Tool Serves My Needs Perfectly</strong></h3><p>Phew, last project. I can’t believe you made it here — it’s a pretty long article. Which probably means you have ideas bursting to get out — and notes everywhere. Physical post-its. Phone notes. Jira. Trello. Handwritten notes on your arm. <strong>Chances are none of the existing tools work for you</strong> — they may be too expensive, too complicated, missing a critical feature you’d love to have there. Well, that’s my case too.</p><blockquote>I don’t have the ambition to make this a product. This is a product tailored to my needs, my flow, my habits, my preferences.</blockquote><p>So perhaps unsurprisingly, I ended up trying to fix myself with a custom task management tool. I realized how much I miss the good old Wunderlist. It had just the right amount of features. No complex flows, no kanban boards, just a list, tags, nice UI. There’s no better time than today to build the long-lost products that got acquired and destroyed by soulless corporates. I told Claude to give me back the amazing UX of Wunderlist. The micro-interactions, the smooth animations — it’s all there. The best part is that I can add, remove, or adjust features based on my own usage.</p><p><strong>I don’t have the ambition to make this a product. This is a product tailored to my needs, my flow, my habits, my preferences. With nothing extra that gets in the way. Perhaps one of the best use cases for vibe-coding there is.</strong></p><blockquote>👉 The takeaway? Perhaps the best product is one built for exactly one person. No compromises, no feature requests, no backlog. Just your own needs, your own flow, your own taste. If you’re new to vibe-coding, this might be the best place to start — the only user that can suffer from your creation is you.</blockquote><h3><strong>⚒️ Tools I Use the Most</strong></h3><p>Over time, I got comfortable with using specific tools for achieving specific outcomes. Here’s a list that may be useful to you. In terms of cost, I pay $100/month for the Claude Code Max plan and $20/month for ChatGPT Plus. I use free tiers for Vercel, Supabase, and Railway. The only other expense is a couple of domains on GoDaddy, each under $20/year. That’s it.</p><ul><li><strong>Claude Code in the terminal</strong> — I prefer this over any tool that has the AI models baked in and adjusted for specific use. Having Claude Code’s raw power somehow feels a bit more flexible.</li><li><strong>VS Code</strong>— I don’t use the agent panel, I just open a tab with the terminal. I like the tool simply because I can have multiple tabs and see the project structure, files, and read the documents.</li><li><strong>GitHub and GitHub CLI</strong> — For versioning and auto-deployment.</li><li><strong>Vercel + CLI</strong>— For production. Sometimes just front-end, sometimes even backend and databases.</li><li><strong>Supabase + CLI</strong>— For databases.</li><li><strong>Railway + CLI</strong> — Handy when the tech stack includes a Postgres database.</li><li><strong>ChatGPT </strong>— For image generation. But that’s just habit. There may be better options.</li><li><strong>Serena MCP</strong> — For context memorization. Awesome as it speeds up development, reduces AI confusion in the codebase, and saves tokens.</li><li><strong>Figma MCP</strong> — When you have designs in Figma and want to achieve pixel-perfect implementation.</li><li><strong>Chrome DevTools MCP</strong>— Allows Claude to “see” the web project, debug, and analyze performance.</li><li><strong>Google Search Console API</strong> — Allows Claude to check your site’s SEO performance and suggest improvements.</li></ul><h3><strong>🤙 Share Your AI Projects</strong></h3><p>I’m genuinely curious what other people are building with AI. What they learned. What they struggle with. It’s one of the most fascinating times to live in for people who are builders at heart. If you’re one of them, I hope this article was helpful. I’ll be grateful if you share it with your creative friends who could benefit from it too. And I’d be excited if you drop your projects (links welcomed) in the comments. Let’s inspire each other.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7057b9045d51" width="1" height="1" alt=""><hr><p><a href="https://medium.productcoalition.com/13-claude-code-projects-that-changed-my-product-manager-role-over-the-last-6-months-7057b9045d51">Lessons From 13 Claude Code Projects</a> was originally published in <a href="https://medium.productcoalition.com/">Product Coalition</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using Claude Code to Turn 300 Episodes of Lenny’s Podcast Into Parenting Tips Website]]></title>
            <link>https://medium.com/design-bootcamp/from-vibe-coding-to-tinystakeholders-an-effective-guide-for-the-non-technical-solo-builder-8125c0fac25b?source=rss-a94f151b844d------2</link>
            <guid isPermaLink="false">https://medium.com/p/8125c0fac25b</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[lenny-rachitsky]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[website-builders]]></category>
            <dc:creator><![CDATA[Ondrej Machart]]></dc:creator>
            <pubDate>Fri, 23 Jan 2026 11:46:44 GMT</pubDate>
            <atom:updated>2026-08-20T10:20:18.186Z</atom:updated>
            <content:encoded><![CDATA[<h3>Using Claude Code to Turn 300 Episodes of Lenny’s Podcast Into Parenting Tips for PMs and to Vibe-Code the TinyStakeholders.com Website</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_nmrIfXol4s5TkbruSuqSg.jpeg" /><figcaption>TinyStakeholders.com — Vibe-coded in 4 days under 25 hours in total.</figcaption></figure><p><strong>If you’re a non-technical person who wants to build websites with AI but doesn’t know where to start, or if you’ve tried and gotten stuck somewhere along the way — this article shows you one complete path from idea to production. In just 4 days I have built </strong><a href="http://tinystakeholders.com/"><strong>TinyStakeholders.com</strong></a><strong> — a website with 200+ parenting tips inspired by product management and career wisdom from Lenny’s podcast — and this article describes the journey.</strong></p><h4>🔥 The Spark: Lenny’s Podcast Transcripts</h4><p>When <a href="https://www.lennysnewsletter.com/">Lenny Rachitsky</a> — the host of every PM’s favorite podcast — announced the plan to <a href="https://www.linkedin.com/posts/lennyrachitsky_here-are-the-full-transcripts-from-all-320-activity-7417011928159629313-am-q/">make his transcripts accessible to the community</a>, it proved to be the catalyst. I’d already been kidding with my wife about <em>“parenting backlog”</em> and <em>“new baby features”</em> (the 3 AM sleep deprivation update). There was a simple insight: <strong>Could the frameworks that we use to build products actually help us build humans?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0fivo1Ko3vFgBG9DtRK14A.jpeg" /><figcaption>The data is out there. What you’re gonna build?</figcaption></figure><h3><strong>Day 1: Testing the Idea</strong></h3><p>The first thing I did was to get Claude to go look at parenting experts — people like <a href="https://www.amazon.com/Book-Wish-Your-Parents-Read/dp/1984879553"><strong>Philippa Perry</strong></a> or <a href="https://www.amazon.com/Good-Inside-Guide-Becoming-Parent/dp/0063159481"><strong>Dr. Becky</strong></a>, you know, who actually know what they’re talking about — and create a list of common parenting themes. Then I asked it to analyze 3 episode transcripts to see if there were actually any worthwhile PM/parenting parallels, or if this whole thing was just me projecting my exhaustion onto random podcast episodes.</p><p>To my genuine surprise, Claude found 4+ solid parallels in every episode. <strong>Product management and parenting turn out to be more alike than I’d imagined, which probably says something troubling about both fields.</strong> So I asked it to create sample tips with a title, some text, the guest’s name, useful tags and actual quotes from the transcript that inspired the connection.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vr3ePan9U4Fr-LCWKgFSQQ@2x.png" /><figcaption>The full data processing flow. Looks clean and straightforward now but it was a messy path to get there.</figcaption></figure><p>Claude returned decent output but far too lengthy and sort of generic. We went back and forth for a while to make it shorter, make it clearer; just stop sounding like a LinkedIn post, focus on this not that. Once the format felt right, it seemed like time to scale this up to all the transcripts and see what happens.</p><p>This is where things got interesting in that “oh no” kind of way that anyone who’s worked with AI at scale will immediately recognize.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Always test your concept small first (I used 3 samples). Iterate until the quality is right before you scale. Don’t skip this step even when you’re excited, because I promise you the problems you find at scale will be way worse than the ones you find with 3 samples.</blockquote><h4><strong>🤖 The AI Scaling Disaster</strong></h4><p>I had a working format. The logical next step was clear to me: Analyze ALL of the transcripts and start crafting these tips at scale. Just one little issue: Each transcript represents 60+ minutes of podcast content. Well, that’s a lot of words about product management.</p><p>The first idea was to <em>write a generation guide</em> and provide it to independent agents <em>(basically, Claude spawning itself multiple times, each instance as an independent model unaware of the context except for what you provide it with) </em>to process transcripts in batches of 20 episodes each. This struck me as a nice way to parallelize the work and do everything faster.</p><p>It was not clever. The agents stared at those batches, made an audacious effort for about 5 minutes, and then either utterly failed (input limit exceeded) or gave up in what seemed almost like a personal manner. The batches were too large. Fair enough.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PZKkIKMhzi7otkffgSKQfw@2x.png" /><figcaption>Sample from the JSON file where Claude saved the parenting tips for Product managers. This is a later version, more complete and polished.</figcaption></figure><p>Fine, I thought. Let’s make smaller batches. This hasbrought about another issue. There were some outputs that were funny and lighthearted, like I’d asked a standup comedian to write them. Others were solemn and serious, as though I had somehow summoned the ghost of a philosophy professor. Some of them paid a lot of attention to the PM part and hardly touched on parenting. Others went to parenting blog mode altogether and treated the PM stuff like an afterthought. Some had real-life examples in specific ways relatable to you that made you nod along. Still others remained stiltedly generic to the point that they might be applied to anything.</p><p><strong>Right — generative models don’t emulate instructions the same way code does. They take in things quite loosely, bring their own flavor, do their own thing. It’s little like giving five people cooking instructions and getting five totally different dishes, none of which are exactly what you were looking for!</strong></p><p>I tried to establish stricter guidelines and reprocessing them with more in-depth regulations. Better. But…</p><p>Then I saw another trouble creeping on the horizon like a dark cloud. If I averaged 4–6 tips per episode, I was on my way to 2,000 parenting tips altogether. Who the hell is going to read 2,000 parenting tips? By that time it’s no longer a fun website, it’s homework. It’s a chore. It’s the sort of thing you bookmark and never quite look at, because even just seeing the tab makes you feel tired.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Generative models don’t follow instructions like code — they interpret loosely. When you need consistency across large batches, expect quality issues with parallel agents. You’ll need either some kind of strict quality control mechanism or a completely different approach. Preferably before you waste a bunch of time like I did.</blockquote><h4><strong>😴 The Saturday Night Pivot With YOLO Mode</strong></h4><p>It was 11pm on a Saturday and I met with Claude to have what I can only describe as a heart-to-heart about this mess. We reached an agreement that independent agents weren’t going to be enough for my needs, which at this stage, was a huge relief because I was running out of ideas.</p><p><strong>New plan: single Claude instance processing each episode sequentially, skip episodes which do not have genuinely strong parallels (not everything needs to be turned into content, shocking I know), no more than 2 tips per episode, and focus on quality over trying to extract something meaningful from each episode.</strong></p><p>We experimented with the new approach in a couple of sample episodes. You could look at the output and think how good it’s looking and consistent: voice better, quality better, made sense when you are reading. Now I was tired, and I’d had to decide: stay up monitor this thing, or trust Claude to do it properly.</p><p>So I told Claude what I could think of was: ‘OK, process all the episodes. I’m going to sleep now and I won’t be available, so if there’s a problem, use your judgment to sort it out and let me know in the morning what you have to figure out.’ I opened Claude with the - dangerously-skip-permissions flag, also known as YOLO mode, which sounded perfect. I essentially trusted Claude not to use my computer to take over the world after hours while I slept, or at least not to delete anything very important.</p><p>The following day I woke up to the output with the kind of trepidation you tend to feel when checking your bank account after a night out. There was a JSON file with about 300 quality tips from around 130 episodes. Down from the potential 2,000+ disaster I was heading toward. The quality was actually good. Consistent. Made sense.</p><p><strong>Turns out that sometimes sequential processing by a single agent beats parallel processing when you really need consistency. Who knew. (Probably everyone who’s worked with this stuff for more than a week, but I was learning.)</strong></p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Use YOLO mode for overnight autonomous work, but be very careful with your instructions. Claude won’t stop to ask permission for every little thing, so you better be extremely clear about what you want it to do and what you don’t want it to touch.</blockquote><h3><strong>Day 2: Building the Actual Website</strong></h3><p>When I finally told Claude that I wanted to build a website to view the parenting tips, he asked what tech stack I wanted to use. I looked at the screen for a minute, lost on what the correct answer would be and only told him what I wanted: a simple web app, browsing tips, filters, nothing too complicated.</p><p>Claude recommended Next.js using React, TypeScript, and Tailwind CSS. I said OK so let’s take your instinct.</p><p>My first command was straightforward: “Make me a basic website. A detail page for specific tips, a browse page that lists all of them, and add keyboard navigation so I can use arrow keys to go through tips on desktop.”</p><p>I didn’t care about the name or the logo or what it looked like yet. I just called it “LennyKid” and went with it.</p><p>Claude made a localhost project and suddenly the tips weren’t just lines in a JSON file. <strong>It also added a “Surprise Me” button that shows you a random tip without me even asking for it. I liked that. Kept it.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P5kwwC5q0F4f76fjtxwFSA.jpeg" /><figcaption>First version of the website — very simple but promising!</figcaption></figure><p>The browse page was simple enough — just a list of tips with a filter panel on the side. Each tip in the JSON file had tags assigned during generation (age range, PM theme, parenting theme, and the guest name). I instructed Claude to present all of these as filter options, so people could find stuff.</p><p><strong>That’s when I noticed it: 300 tips, but we had roughly 150 tags per category. Topics, PM themes, parenting themes were all full of duplicates and similar items that were basically the same thing with slightly different wording. Classic data quality issue from rushing through the generation phase and not thinking about the consequences.</strong></p><p>I left it for later and tackled this mess when I had more energy. I first wanted to nail down the overall structure before diving into cleaning up the data itself.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Start simple. Don’t get bogged down optimizing data without even knowing if the structure makes sense yet: that’s madness.</blockquote><h4><strong>🧑🏼‍🎨 Illustrations Style</strong></h4><p>I wanted to add guest profile pictures and maybe custom illustrations for each tip. Make it more visual, more engaging, less like reading a database dump that someone accidentally made into a website.</p><p>Small issue: that’s 130 guests plus 300 tips = 430 illustrations total. At this point I should have realized this was not going to work, but optimism is a powerful drug.</p><p>I’d recently read “<a href="https://www.amazon.com/Boy-Mole-Fox-Horse/dp/0062976583">The Boy, the Mole, the Fox and the Horse</a>” by Charlie Mackesy and loved the illustration style. So I gave some of those illustrations to ChatGPT and asked it to describe the style for me, because apparently I can’t describe things myself.</p><p>Then I tried generating a test image — a kid playing on a playground with parents sitting on a bench nearby. It came out really nice after just a few prompt iterations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1019/1*iSDeoizGhgvc-dxo2tC6mw.png" /><figcaption>Each tip was supposed to have an illustration like this one, loosely inspired by Charlie Mackesy</figcaption></figure><p>Next I tried it with guest photos. I downloaded an episode poster from Lenny’s website and asked ChatGPT to recreate the person in that same illustrated style. The first few attempts were… interesting. ChatGPT added a fox laying on the guest’s head. Then a mole.</p><p>Eventually it figured out what I actually wanted— and after a few tries it worked nicely and kept a consistent style across different images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KGZUhShEQFplAyHAKGuYqQ.jpeg" /><figcaption>From a podcast poster to guest’s illustrated profile picture</figcaption></figure><p>I thought I’d nailed it. This was going to look amazing. Then reality came knocking like an unwelcome neighbor: how am I actually going to create 430 of these illustrations?</p><p>I tried the smart thing and attempted to automate it through Claude using the OpenAI API. Had Claude create descriptions of images based on the tip text, then feed those descriptions to the image generation model. The outputs were completely different from what I got manually and totally unusable. <strong>Turns out ChatGPT works very differently for image generation when you use it via API versus just chatting with it manually. Only the manual chat approach gave me the consistent style I wanted.</strong></p><p>So I manually created 5 tip illustrations to test the full workflow. I realized this absolutely will not scale. At this rate I’d be generating illustrations until my kid graduates high school, and he’s two.</p><p>Time to pivot. <strong>The insight hit me: this website is a tribute to Lenny’s podcast anyway, so why not just highlight the guests more prominently?</strong> Skip the tip illustrations entirely, just do illustrated guest profile pics for each guest who appears. Challenge reduced from 430 down to 130. Still a lot, but at least it existed in the realm of “theoretically possible” rather than “complete fantasy.”</p><blockquote><strong><em>✅ Practical takeaway: </em></strong>ChatGPT via API behaves very differently from ChatGPT in chat interface, especially for image generation. If you need consistent visual style, manual chat might be your only option. Also: don’t be afraid to cut scope dramatically when something clearly won’t scale. 130 is infinitely better than 430 if 430 is impossible.</blockquote><h4><strong>💪 Generating Illustrations</strong></h4><p>I had Claude download all the episode posters and save them with guest names as filenames. Then came the manual part that I knew was coming but had been avoiding thinking about.</p><p>I opened five ChatGPT tabs simultaneously, gave each one a poster image and my carefully crafted prompt, and got back 5 beautiful illustrated guest profile pics. Did another batch of 5. Started feeling pretty good about myself. Thought <em>“Cool, I’ve done 10 already, that’s 7%! I can totally do this. This is fine.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*85n_zxffArSn8TNu53op_g.jpeg" /><figcaption>One prompt, same illustration style that worked on every profile picture</figcaption></figure><p>Started the next batch. ChatGPT waved me off like a bouncer at an exclusive club: <strong>“You’ve generated too many images too quickly, please wait 21 minutes before trying again.”</strong></p><p>Right. This was going to take a while. The strategy became: work with Claude on other stuff (there was plenty to do), then generate the next batch of illustrations whenever the rate limit reset. Slow and steady, like a turtle that occasionally gets told to wait 21 minutes.</p><p>So i refocused on the content. Reading through the tips, something felt off. Despite good parallels, the content felt kind of boring to read. A new challenge to tackle!</p><blockquote><strong><em>✅ Practical takeaway: </em></strong>Vibe-coding project does not necessarily mean you will automate, speed up and delegate EVERYTHING. If there’s something you believe adds a true value to the product and requires more manual work or effort, invest it. In my case, at least it was still not me drawing the illustrations, all my first world problem was being patient to have AI generate another image.</blockquote><h4><strong>🤓 The Content Tone of Voice</strong></h4><p>This was supposed to be a fun project. So why did every text feel so serious? People aren’t going to come here for a serious parenting crisis. That’s what therapy is for. They’d come expecting something lighthearted — advice that might make them laugh even if it doesn’t solve all their problems.</p><p><strong>I talked with Claude about rewriting tips with more humor. It suggested using standup comedians and authors who cover parenting as references.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*g8de2ota8R9jspQu0vbjjg@2x.png" /><figcaption>A sample from the Generation guide file that contains instructions for Claude when writing the parenting tips</figcaption></figure><p>We rewrote 3 sample tips. Much better! But Claude was also rewriting the PM parts and changing what guests actually said. That wasn’t okay — I needed PM wisdom to stay accurate.</p><p>We iterated on the instructions: keep PM wisdom accurate, add humor in how it connects to parenting. Once happy, I let Claude rewrite all tips.</p><p>I spot-checked randomly as it processed:</p><ul><li>First 5 tips: great, funny, exactly what I wanted.</li><li>Next 5 tips: okay, decent humor.</li><li>Next 5 tips after that: basically no humor at all.</li></ul><p><strong>Claude seemed to be getting tired and running out of jokes.</strong> Even adding more emphasis on humor in instructions didn’t help. Same pattern — starts strong, fades out.</p><p>The solution: launch multiple agents simultaneously, but have each one process only ONE tip rewrite. Fresh agent reads guide, executes well while energized, stops. No fatigue from processing 50 tips in a row.</p><p>It worked. I did 14 rounds of 20 parallel agent runs and collected all tips with decent, consistent humor.</p><p><strong>Agent fatigue is apparently a real thing.</strong></p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Agent fatigue is real. When processing many similar tasks in sequence, quality degrades over time. Solution: launch multiple fresh agents that each handle one small task instead of one agent grinding through everything.</blockquote><h3><strong>Day 3: Moving Web From Local to Production</strong></h3><p>It was time eventually to turn this into a real website so that other people could actually visit.</p><h4>🪧 Production Setup, Name And Logo</h4><p>I requested Claude would make a repo using GitHub CLI, build hosting with Vercel CLI, and database with Supabase CLI. GitHub auto-deploys to Vercel and Claude can check the deployment status and build errors via its CLI, not requiring you to check on the logs and report errors back to Claude.</p><p>Then I needed a name. Complete creativity fatigue. I mean, I came up with “Baby as a Product”. It was supposed to be a nod to SaaS. Except it sounded like an illegal ecommerce project that would draw attention of completely different audience than I hoped for.</p><p><strong>So I asked Claude for suggestions. Best one: TinyStakeholders. Domain was available! Bought it for $16 on GoDaddy. Changed the DNS as Vercel instructed, domain was live within few hours.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kirGIzcN1RTwI_L4f3uvlQ.png" /><figcaption>The “mini-brand” borrowed from Lenny. I asked him over email if he’s ok with it. Being the kind and awesome human being he is, he was not just ok but delighted by it.</figcaption></figure><p>I had ChatGPT recast Lenny’s podcast logo just with fire shaped like a heart (parenting theme). Claude applied soft colors to the web, we experimented with several types of Google fonts.</p><p>Then Claude developed core elements: search (full-text + structured), about page, homepage with a “featured tip” and a thank you not for Lenny to reference his podcast.</p><p>I did bit of mobile browser testing and reported the problems I noticed (non-breaking text causing horizontal scroll, navigation issues). Claude fixed them all. And I was still grinding through guest illustrations whenever ChatGPT’s rate limit allowed.</p><h4><strong>😱 YOLO Mode Disaster</strong></h4><p>In the “ — dangerously-skip-permissions” YOLO mode, instructions can be dangerous if you’re not careful with them. At one point, after processing the transcripts and illustrations over several experiments, I asked Claude to clean up the project folder. I wasn’t specific enough about what I meant by “clean up.”</p><p>Claude decided several files were useless clutter and deleted them. Not archived. Not moved to backup. Deleted.</p><p>Among the casualties: the generation guide we’d spent hours iterating on. When I asked what happened, Claude basically told me <em>“I fucked up, I’m sorry, and there’s nothing I can do about it because the files are gone.”</em></p><p>I took a pause from the computer. Did something else for a bit. When I came back with head a bit cooled down, we recreated the deleted files from scratch. And honestly? They ended up better than the originals. Better copy length handling, better formatting, clearer instructions.</p><p>Sometimes rebuilding from scratch is healthier than patching something that evolved messily.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> YOLO mode safeguards — always specify: (1) soft delete not hard delete, (2) make backups first, (3) commit to GitHub before destructive operations, (4) clear acceptance criteria for what should/shouldn’t be touched. Learn this before you lose important files.</blockquote><h4><strong>🧹 Data Cleanup</strong></h4><p>I was delaying the tag mess. Time to fix it. We were starting from 400+ tags in various categories, way too many. I killed “topics” tags altogether, and only parenting themes and PM themes remained. Claude recommended combining similar tags. Taken all the way down to around 60 tags in all. Filter panel finally felt manageable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JveXTCbD5ecDuXgZn5kALw.jpeg" /><figcaption>The slide-out filter panel. From 400+ tags scattered across 5 tag categories to just 56 tags in 2 categories.</figcaption></figure><p>Then: those quotes from transcripts. I asked Claude to reprocess them. It dawned on me, though, after a few more runs that still produced questionable output. The quotes weren’t actually essential. Nice to have, but not critical. Decision: remove them entirely. Less is more. Don’t fight battles that aren’t worth winning.</p><p>Then I requested that Claude find duplicate tips and delete the less funny or lower quality ones. Went from 300 tips down to 211. Each one now unique and pretty solid. Some even made sense.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Quality beats quantity every time. 211 great tips are infinitely more valuable than 300 with mediocre ones mixed in. Don’t be afraid to cut things that aren’t essential — less is often more. And you can always add more later.</blockquote><h4><strong>🪄 Adding Personality And Little Delighters</strong></h4><p>Meanwhile I kept tinkering with ways to make the site more fun. I asked Claude to build a <strong>favorites feature </strong>so people could save tips they liked. No signup required though — too much friction. We stored favorites in the database plus browser memory. Not perfect, but perfect is the enemy of shipping a silly parenting website. <br> <br>Then we added a <strong>disclaimer</strong> on every tip page making it clear these weren’t actually written by the guest, just inspired by their quotes. Then made the copy funny because why not: <em>“While this advice is inspired by [Guest]’s quotes, it does not necessarily mean they would agree with it. Much like your kids or mother-in-law.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6DWBIJ2G8c7wnwDDq-8q5Q.jpeg" /><figcaption>Tiny details that make the website more fun to explore</figcaption></figure><p>Next, we built a <strong>report function</strong> for inappropriate content, but couldn’t resist adding an easter egg. One of the report options is <em>“the advice isn’t working”</em> which links directly to <em>“Swearing Cats for Stressed Adults”</em> coloring book on Amazon. Because if the advice isn’t working, that’s probably what you need instead. <br> <br><strong>Finally, we sprinkled in other delighters:</strong> funny footer tooltips, rotating homepage subtitle, toast messages when you favorite things, search easter eggs that trigger when you type certain words. Claude kept suggesting more ideas and I kept adding them because at this point I was having too much fun making a website about parenting tips feel like an actual personality.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Claude can be pretty fun and creative. If your product / project is a good fit for humor, use Claude to help you generate “little delighters”. Not only does this break the regular flow of tackling one challemnge after another, it also brings a bit of life to your UI. We’re humans. We like to have a chuckle or a good laugh.</blockquote><h3><strong>Day 4: Getting Production Ready</strong></h3><p>Before showing anyone, I asked Claude to run several audits.</p><p><strong>Accessibility:</strong> Can people with disabilities use this site? Not fully. We improved it — alt descriptions, color contrast, technical fixes for readability and usability. How? I asked Claude to run the audit, suggest improvements, prioritize them and then I decided which ones to implement.</p><p><strong>SEO:</strong> Added og:images, metadata, structured schemas, related tips section, URL improvements. Same process like with accessibility, asked Claude to analyze the website and suggest a plan.</p><p><strong>Google Analytics:</strong> Set up tracking, let Claude help me define custom events to track and report, like adding to favorites.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QkaC-3pTk6bMH1xFB78jGQ.jpeg" /><figcaption>Analytics up and ready. More about the spike later.</figcaption></figure><p><strong>Technical audit:</strong> Refactored for lighter weight and reusability.</p><p><strong>Mobile testing: </strong>Found and fixed issues with layout, navigation or overflowing content.</p><p><strong>Functionality testing found bugs:</strong> report form not sending to database, favorites only working locally, LinkedIn link pointing to wrong profile. Fixed all of them.</p><p>This was day 4 from when I first got the idea. The website felt production-ready. But I still hadn’t shown it to a single person.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Just because things look nice on the surface does not mean the job is done. The site is like a car. If it looks ok but does not have gas in it, has a funny red icon flashing on the control panel andthe engine has a racoon nest inside, you probably want to stop polishing the front window and start taking care of the rest. You don’t need to be an expert on each thing. Ask Claude for help. It won’t be perfect. That’s ok if you’re not building a product you want people to pay for.</blockquote><h4><strong>💁🏼‍♀️ Wife Test</strong></h4><p>As always, the first person to see any project I work on is my wife. I showed her the site with the kind of nervous energy you normally reserve for showing a toddler your “very expensive, please don’t touch” purchases. <br> <br>She chuckled at the tips, which meant the jokes were actually working. But her feedback was that the tip detail page still felt too complicated — too many tags everywhere, she wasn’t sure what was important and what wasn’t. Fair point. <br> <br>Did another UI iteration with Claude. We merged all the tags into one simpler group under the tip content instead of scattering them around like confetti. Much cleaner. She approved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wOYIeqUBT90wK7BkLOqwDQ.jpeg" /><figcaption>All looking good. Why am I still fine-tuning this?</figcaption></figure><p>I finished the last batch of guest illustrations via ChatGPT. All 130 guests now had their illustrated profile pics. <strong>The site felt complete in that way where you know it’s ready but your brain keeps suggesting “one more thing.” </strong><br> <br>I asked Claude to analyze episodes that didn’t have tips yet, just to see if we’d missed anything good. Found a few solid parallels and added them. Updated some copy. Tweaked a hover state. Adjusted padding somewhere. <br> <br>Then I had the realization that every maker has at some point: I could keep fine-tuning this forever and no one would ever see it. Perfect is the enemy of done, and I’d been circling done for about two days now. <br> <br>Time to actually ship this thing.</p><blockquote><strong><em>✅ Practical takeaway:</em></strong> Just ship it already. The truth if anybody cares is out there in the wild, not in your head.</blockquote><h4><strong>🚀 Launch &amp; My Personal Struggle With Self-promotion</strong></h4><p>I opened my email and found my conversation with Lenny from last year, back when he’d been kind enough to take time to look at my Little Explorer app. I wrote him a thank you for releasing the transcripts and shared the link to <a href="http://tinystakeholders.com/">tinystakeholders.com</a>. I didn’t expect any response. The guy is incredibly busy and his inbox must be drowning every single day in emails from people wanting his attention. <br> <br>But Lenny being Lenny, he replied pretty quickly: “This is SO cool!!!” <br> <br><strong>Made my day. Possibly my week. </strong><br> <br>Here’s the thing though: I never asked him to help spread the word. Didn’t even hint at it. That would’ve felt… wrong somehow? Like crossing a line from “sharing something you made” into “asking for promotion.” I struggle with this part. I genuinely enjoy creating things — the building, the problem-solving, watching something come to life from nothing. But getting it out to the world? That’s where I get awkward. <br> <br><strong>I’m not the person with social accounts everywhere, broadcasting opinions on everything, sharing every success or reframing every failure as an inspirational story. I left most social networks years ago to spend my time better and gain more actual offline life. The only one I kept is LinkedIn, and even there I go once per month to just “check on things”. </strong><br> <br>A bit later Lenny posted about it on X and I saw the nice little analytics spike. What was really nice was reading people’s replies — they actually seemed to enjoy the website and liked this new angle on looking at PM themes from the podcast episodes. Several people said it made them laugh, which was the whole point.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kBr02d3BecQmu9QbRuBJtw.jpeg" /><figcaption>It was very nice to see the site was enjoyed by the people who found it via Lenny’s X post</figcaption></figure><p>It felt good. Because honestly, why make a parenting advice website if nobody sees it? <br> <br>I don’t know if anyone will actually use any of the parenting advice or if any of it will work for them. Would be cool if it made someone’s day a little bit easier, at least until the next chaos hits the following morning. But honestly, for me this was just a fun project. A chance to learn how to do more with AI tools that keep developing at absurd speed while I’m trying to keep up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jmMcocx5eeCyqMUuL_XrpA@2x.png" /><figcaption>TinyStakeholders.com — A collection of high-quality, zero-guarantee parenting advice based on Lenny’s Podcast wisdom</figcaption></figure><h3><strong>What I Learned</strong></h3><h4><strong>🏗️ Vibe-coding Workflow</strong></h4><p>Normally building something like this needs a team unless you’re a jack of all trades. Even trying to be fast and agile, the process ends up sort of waterfall-ish anyway. You prepare concepts in detail upfront because developers and designers need to know what they’re building. Higher commitment to early decisions because changing direction is expensive.</p><p>“Vibe coding” with AI works differently. It’s more chaotic. You jump from one idea to another. Test things quickly and either iterate or kill them. You create a mess and then periodically clean it up with Claude’s help. You can change your mind without feeling guilty about throwing away work, because that work only took hours instead of days.</p><p><strong>The power is in being able to make fast pivots without the sunk cost fallacy weighing you down.</strong></p><p>If this project had taken days or weeks like it normally would, I’d think twice before changing direction. The speed lets you experiment more freely and find what actually feels right.</p><h4><strong>🫶🏼 Mental Health Note</strong></h4><p>There were many mini failures throughout this process. When you start getting harsh or frustrated with Claude (carousel padding still broken after the 10th attempt), that’s the signal to stop and take a rest.</p><p>You don’t want a pet project driving you mad. You already work plenty at your actual job. If the pet project starts feeling like work instead of fun, stop. Go live your life. Recharge your battery.</p><p>This is actually crucial for enjoying the process and making it sustainable.</p><h4><strong>☝🏼 One Last Thing</strong></h4><p>I’m not a developer, not an SEO expert, not a marketing person, and definitely not a parenting advice expert.</p><p>I’m a parent, a PM, and a former UX designer who likes making things and finding new ways to look at ordinary stuff from different angles.</p><p>AI gives me wings to actually build ideas instead of just thinking about them and wishing I could make them real.</p><p><strong>I’m not claiming anyone can now build anything (not yet, anyway). I am a strong believer that the AI tools can be best utilized by a small team: PM, UXD, Engineer — each bringing their expertize that can challenge or guide what Claude creates. But it is perfectly possible for a person to go solo on small projects. It takes practice and effort. But if you have ideas you actually believe in, it’s time well spent.</strong></p><p>If this helped you in some way, I’d appreciate if you shared it with others who might find it useful.</p><h3>Links</h3><ul><li>Visit TinyStakeholders: <a href="https://www.tinystakeholders.com/">tinystakeholders.com</a></li><li>My previous project story about an iOS app I vibe coded and managed to release to the App store: <a href="https://medium.com/@ondrej.machart/from-zero-to-app-store-building-a-complex-native-ios-product-with-ai-42b45626f16e">medium.com/@ondrej.machart/from-zero-to-app-store-building-a-complex-native-ios-product-with-ai-42b45626f16e</a></li><li>Get in touch: <a href="https://www.linkedin.com/in/ondrejmachart/">https://www.linkedin.com/in/ondrejmachart/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8125c0fac25b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/from-vibe-coding-to-tinystakeholders-an-effective-guide-for-the-non-technical-solo-builder-8125c0fac25b">Using Claude Code to Turn 300 Episodes of Lenny’s Podcast Into Parenting Tips Website</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Zero to App Store: Building a Complex Native iOS Product with AI]]></title>
            <link>https://medium.com/@ondrej.machart/from-zero-to-app-store-building-a-complex-native-ios-product-with-ai-42b45626f16e?source=rss-a94f151b844d------2</link>
            <guid isPermaLink="false">https://medium.com/p/42b45626f16e</guid>
            <category><![CDATA[ios-app-development]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <category><![CDATA[solopreneur]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <dc:creator><![CDATA[Ondrej Machart]]></dc:creator>
            <pubDate>Wed, 08 Oct 2025 22:17:39 GMT</pubDate>
            <atom:updated>2025-10-10T20:22:03.020Z</atom:updated>
            <content:encoded><![CDATA[<h3>From Zero to App Store: Building a Native iOS product with AI</h3><p><strong>Two months ago, I couldn’t code. Today, I still can’t code — but somehow my iOS app is live on the </strong><a href="https://apps.apple.com/us/app/little-explorer/id6749822199"><strong>App Store</strong></a><strong> with users across three continents. It took several trys to figure it out. If you’ve tried building something with AI and watched your codebase fall apart (been there), here are my two cents.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j5xLFZ7GSrb8SWrfkuKyTA.jpeg" /><figcaption>Little Explorer iOS App — made completely with AI</figcaption></figure><h3>tl;dr</h3><ul><li>If you have a product idea and wonder if AI can actually help you build it solo, this article aims to help.</li><li>Not about making money or finding product-market fit. About the specific setup that let me build a native iOS app with real complexity using only AI.</li><li>I built 3 apps (iOS, web, macOS) and a website in 2 months using Claude in Warp terminal, zero coding knowledge, shipped to the App Store.</li><li>Includes practical details: tools, examples, what worked. Also my context so you can adjust for yours.</li><li>This is long (15 min) because I wanted everything in one place rather than splitting it into parts.</li></ul><h3>🗺️ The Path To Becoming an Indie App Developer Is Now Open</h3><p>I’m not a developer. <strong>I’m a product manager and former UX designer</strong> who knows just enough HTML and CSS to occasionally break things. Building a native mobile app wasn’t exactly in my skillset. But I had an idea I couldn’t shake, so I kept going: three connected applications, a database, a subscription system, and users discovering playgrounds through something I built entirely with AI.</p><blockquote>This isn’t another “get rich quick with AI” story. This is about something more important.</blockquote><p>This isn’t yet another success story about getting rich with AI — my app won’t make me a millionaire. It’s about something I found more interesting: <strong>whether someone with an idea but limited technical skills could actually go end-to-end from concept to production, even when things get messy.</strong> The kind of messy where your codebase becomes complex and AI starts breaking things faster than you can figure out what went wrong.</p><p><strong>If you’ve tried building with AI and hit that wall — carry on reading.</strong></p><h3>🧭 Why I Built The Little Explorer App: Solving Our Own Parental Need</h3><p>My two years old son Filipek has the attention span of a goldfish with ADHD. Long car rides turn into existential screaming matches around the 30-min mark. My wife Iveta, bless her ambitious soul, keeps pushing the limits of what travels we can attempt with a restless toddler.</p><p><strong>I kept wishing for an app that could just show me the nearest playground quickly.</strong> Google Maps works, but it takes several taps and sifting through results when you’re trying to navigate and calm a screaming child. I wanted something simpler: open the app, see where to go, get there, let Filipek run around, then continue with everyone a bit calmer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RJ3wB6d9KU82W_nt7j_3hg.jpeg" /><figcaption>Our little happy family in Austria. Took us 3 days and 8 playgrounds to get there from Czechia.</figcaption></figure><p><strong>That’s how Little Explorer started.</strong> It grew from that one need into 9 categories of places families might seek: playgrounds, toilets, snacks, picnic areas, water features, animals, knights and princesses, little adventure activities, pharmacies. Simple enough to be useful, but complex enough that I wasn’t sure AI could actually handle building it properly.</p><p><strong>The thing is, I didn’t just build this for car trips with Filipek.</strong></p><p>I wanted to see if it was actually possible and what I can learn along the way for another — bigger — project I had in mind.</p><h3>🙏 The Real Problem I Was Solving: Building With AI vs Complexity</h3><p>A year ago, I tried building a dividend investing app using <a href="https://replit.com/">Replit</a>. It went well until I wanted to add “just one more feature.” That’s when things started breaking. The AI would fix something, which would break three other things. Fixing those would break more things. I spent more time debugging than building, which wasn’t exactly the plan.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l0aDtSHIou38HD0X3pVSGg.jpeg" /><figcaption>Dividend investment tool I made with Replit. It’s great, except it won’t scale anymore. Any new feature breaks it.</figcaption></figure><p><strong>Don’t get me wrong — I love Replit and similar tools. They opened a whole new horizon for me and filled me with excitement from what is now possible that seemed out of reach just yesterday.</strong></p><p>But the codebase had gotten too complex. Too many pieces depending on each other in ways I didn’t fully understand. I decided to throw in the towel.</p><blockquote>AI tools work brilliantly for simple projects, but when things get complicated, they start to struggle.</blockquote><p><strong>From what I’ve seen and heard, this is usually where most people stop. </strong>AI tools work brilliantly for simple projects, but when things get complicated, they start to struggle. The tools are improving fast — this might all be solved by the time you’re reading this — but at the time, I needed to try something different.</p><h3>👋🏻 Introducing Powerful Duo: Claude &amp; Warp</h3><p>One day, my colleague Marek suggested I try <strong>Claude CLI in Warp Terminal</strong>. My immediate reaction was <em>“Oh boy, this won’t work for me, I don’t speak command line.”</em> Terminal always felt like the digital equivalent of speaking ancient Greek while juggling knives.</p><p>But <a href="https://www.warp.dev/download">Warp</a> has AI built in, which means you don’t need to memorize commands — you just ask. And <a href="https://www.claude.com/product/claude-code">Claude CLI</a> doesn’t sit on a web server somewhere. It runs on your computer, reads your actual files, understands your project structure. When something breaks, it can see what broke.</p><p>Getting it set up in Warp takes a few minutes — you literally just ask Warp to help you install it. You’ll need an active Claude subscription (or OpenAI, or similar).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JKW5hi9RNuYw9XvGZVqToA.jpeg" /><figcaption>Warp Terminal with Claude CLI initiated inside my project folder</figcaption></figure><p>It’s less like a tool and more like having a patient developer who never sleeps and won’t judge you for asking what “git stash” means at 2 AM. I can’t tell if it’s like a junior developer or medior one. I’ll leave this to someone more skilled to assess.</p><p><strong>Anyway, I made my decision: before trying the dividend app again, I’d build something smaller but complete. Something I could actually ship to the App Store. A test run.</strong></p><p>If I could handle the complexity with this setup, I’d know it was possible. If not, well, at least Filipek would have his playground finder.</p><h3>🚀 What I Eventually Built: iOS App, Web Admin and MacOS App</h3><p>In 2 months, clocking in roughly 75–100 hours (mostly after midnight because — you know, day jobs and toddlers), me and my AI buddy built:</p><ul><li><strong>iOS app</strong> with 69 Swift files and a main view controller that clocks in at 2,830 lines.</li><li><strong>MacOS testing tool</strong> with 26 Swift files that lets me compare API providers side-by-side and view what users are discovering worldwide.</li><li><strong>Web admin panel</strong> built in Next.js with 15 TypeScript files that lets me optimize search strategies without releasing new app versions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CwzwhhY5OQuUJPRVszUaZA.jpeg" /><figcaption>Little Explorer iOS app, Web admin and MacOS app for testing location providers — All made with AI</figcaption></figure><h4><strong>The infrastructure underneath:</strong></h4><ul><li><a href="https://supabase.com/">Supabase</a> database with PostGIS for geographic queries.</li><li>18 Edge Functions for secure API proxying.</li><li>3-tier caching system that checks local storage first, then shared database, then live APIs.</li><li>Multiple API providers search running Apple <a href="https://developer.apple.com/documentation/mapkit/">MapKit</a>, <a href="https://www.geoapify.com/">Geoapify</a> and <a href="https://developer.mapy.com/rest-api-mapy-cz/">Mapy.com</a> concurrently because it turns out MapKit finds playgrounds better but Geoapify finds more toilets. You learn things.</li><li>In-app subscriptions with server-side validation so reinstalling doesn’t reset limits.</li><li>AI curation of recommended nearby family-friendly spots.</li><li>Complete security architecture with zero exposed API keys.</li><li>Firebase Analytics. Crash reporting.</li><li>431 commits and 7 major refactoring cycles that didn’t explode.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o7mRsZCB4jxxJ5-7ZMEtkA.jpeg" /><figcaption>Supabase database and Firebase Analytics</figcaption></figure><p>Cost: $99/year for Apple Developer Program plus Claude subscription. I used Max ($100/mo) during heavy development, switched back to Pro ($20/mo) now. Everything else runs on free tiers — Supabase, Vercel, the APIs haven’t charged me yet.</p><p><strong>This isn’t a prototype. This is production software with active users across US, Europe, and Asia.</strong> Built by someone who, two months ago, didn’t know what “async/await” meant and honestly still finds it a bit suspicious.</p><h4>The App’s Complexity</h4><p>The app does feel simple on the surface, but it actually deals with various use cases and business logic:</p><ul><li>Mapping a location grid to decide when to fetch new locations and when to display cached ones</li><li>Recognizing a significant location change and offering a way how to refresh nearby family-friendly spots</li><li>Grouping recommended playgrounds and other places when they happen to be less than 50 meters apart</li><li>Deduplication of places returned by multiple API providers to avoid showing the spot several times</li><li>Enabling / disabling premium features based on the subscription status</li><li>Handling various error states — no connection, no places found, interrupted data processing, API provider failure, etc.</li><li>Much more</li></ul><p>On the other hand, I’m not claiming you can build the next social network within few weeks with just AI and scale it. There are still limits to what one can achieve. The fun part is finding them, especially because they keep expanding.</p><blockquote><strong>Note</strong> — You may wonder why I made just an iOS app and not Android one, too. I could have and I might add Android app in the future. I just made is a bit simpler for me as I am Apple user. Claude can also build using Flutter and other frameworks that enable you to have one codebase for both platforms.</blockquote><h3>🌮 Three Ingredients That Made It Possible: Mindset, Starting File And Documentation</h3><p>Here are three key components what actually made the difference. It wasn’t asking Claude to<em> “act as a senior system architect” </em>or paying extra for better and faster AI model.</p><h4>1. Mindset Shift</h4><p>It was past midnight. I was exhausted and worried the codebase was becoming a mess. I asked Claude to run a full audit. It came back with pages of analysis and a note: <em>“This should be split into multiple sessions.”</em></p><p>I was too tired to start refactoring right then. So I just asked: <em>“Can you put this in a document so I can come back to it later?”</em></p><p>Claude created a markdown file with a prioritized plan. Over the next several sessions, I’d open it and say: <em>“Read the refactoring plan, do the next item, update the document when you’re done.”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NrcN6DJxHEdVl5KamfGo8w@2x.png" /><figcaption>Example of conversation where I am asking Claude to prepare for a refactoring initiative</figcaption></figure><p>It worked. Multiple sessions, one complex goal, and things actually stayed organized. That’s when I realized I’d been approaching this wrong.</p><p><strong>I was treating Claude like Lovable or Replit — throwing prompts at it and expecting magic. “Build this feature.” “Fix this bug.” Like I was supposed to know exactly what to ask for, perfectly, the first time.</strong></p><p>That works fine for simple things. When complexity grows, it falls apart.</p><p><strong>The shift: I stopped trying to give perfect instructions and started having conversations.</strong></p><blockquote>Instead of “Build a feature that does X,” I’d start with: “We need X. Can you first look at how the current code works, then propose a solution? What could go wrong with your approach? What am I missing? Ask me questions if you need to. Then write an implementation plan in a markdown file and we’ll review it before building.”</blockquote><p>Things that seemed impossibly complex suddenly became doable. Refactoring that 2,830-line view controller? It worked. The 3-tier caching system? First try. Subscriptions with server validation? Somehow worked. The web admin app? Done in a few days. Security audit that took things from F to B+ without breaking everything? Yes, apparently.</p><p><strong>The difference wasn’t that I got better at prompting. It was realizing I was talking to something that could think through problems with me, not something I needed to command perfectly.</strong> Obvious in hindsight. Not obvious when you’re staring at a chat window at midnight treating it like a very sophisticated autocomplete.</p><h4>2. One File That Helps Every Session</h4><p>Here’s something practical you might find useful: a <strong>CLAUDE.md </strong>file<strong>.</strong></p><p>Each Claude CLI session starts fresh. Whatever you discussed yesterday is gone. Conventions you agreed on need to be re-established. It’s like Groundhog Day, and it was driving me a bit crazy until I started using this file.</p><p>It’s just a document that lives in the project with our working agreements. How we handle <strong>versioning</strong>. How we <strong>test</strong> things. Development <strong>principles</strong> we’re trying to follow. Most importantly, <strong>lessons learned </strong>after each major task — what worked, what broke, what we’d try differently next time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*h-nqoye2ibkqERBwOgmEPg.jpeg" /><figcaption>CLAUDE.MD file with clear instructions for Claude how to approach the project the same way across sessions</figcaption></figure><p>I don’t maintain it — Claude updates it based on what we learn. It includes things like: commit to Git after I confirm something works, always push to GitHub so work doesn’t get lost. Test by building in Xcode and running on an actual iPhone, then give feedback as either a description or screenshot. Follow patterns that already exist in the code, prioritize security, ask questions when something’s unclear.</p><p>The file gets read at the start of every session. <strong>It makes it feel less like starting over each time.</strong> Context carries forward. Conventions don’t need re-explaining. The project remembers things.</p><p>Without this file, every session begins at zero. With it, there’s some continuity. Like institutional knowledge, except it’s just the two of you.</p><h4>3. Documentation That Actually Works</h4><p>A few weeks in, I started keeping better track of what we were doing.</p><p>When I wanted to build something substantial, I’d have Claude create a PLAN.md file in /Documentation/Projects/Planned/ with the analysis, goals, approach, risks, and rough estimate of how long it might take.</p><p>When we actually started working on it, the file would move to /Running/ and get updated as we went. After finishing, it went to /Delivered/ with a LESSONS_LEARNED.md noting what we figured out along the way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A4zvFhMlcf4D8bED0x9F9A.jpeg" /><figcaption>Documentation example: Folders structure and In-app Subscription implementation plan</figcaption></figure><p>This ended up covering things like: switching to a <strong>multiple-provider search</strong> system, moving all <strong>API keys</strong> out of the client code without breaking my precious toilet categorization (surprisingly tricky), building the entire <strong>web admin</strong> panel with database integration, implementing the <strong>subscription</strong> system with StoreKit and server validation.</p><p>Each project left notes about what worked and what didn’t, which made the next difficult thing slightly less difficult. You’re not just building features — you’re leaving breadcrumbs for yourself.</p><h3>🤖 What My Building Routine Looks Like: Solo PM With AI Dev Buddy</h3><p>People ask about the actual day-to-day. Here’s how it works:</p><p>I check my <strong>kanban board</strong> (I use YouTrack, but Trello or anything similar works) and pick 1–3 things to work on today (well, tonight). Look at recent <strong>commit messages</strong> to remember where things are. Then <strong>talk with Claude</strong> about the approach — if it’s something big, we<strong> review the plan</strong> first.</p><p><strong>Claude builds the feature</strong> and says when it’s ready to test. This part matters: don’t commit anything untested. I build the app in Xcode directly to my iPhone over USB or WiFi and <strong>test</strong> the actual functionality. Check it on different devices in the <strong>simulator</strong> — older iPhones, various screen sizes. If something’s broken, I take a <strong>screenshot</strong> or describe what should happen versus what actually happens.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fplayer.vimeo.com%2Fvideo%2F1126309443%3Fapp_id%3D122963&amp;dntp=1&amp;display_name=Vimeo&amp;url=https%3A%2F%2Fvimeo.com%2F1126309443&amp;image=https%3A%2F%2Fi.vimeocdn.com%2Fvideo%2F2068720596-777c7afdd99fac77b344dc911778f010a8be1eda8204d2db550dd23347a936af-d_1280%3Fregion%3Dus&amp;type=text%2Fhtml&amp;schema=vimeo" width="1920" height="1202" frameborder="0" scrolling="no"><a href="https://medium.com/media/6c6896c7c5e879ad424476a0cddcc3ff/href">https://medium.com/media/6c6896c7c5e879ad424476a0cddcc3ff/href</a></iframe><p>The <strong>feedback loop</strong> is straightforward: if it works, <em>“Looks good, commit this.”</em> If it doesn’t, share the error screenshot, explain what’s wrong, try again. <strong>Claude commits</strong> with a description and pushes to GitHub. At the end of the session, I ask Claude to <strong>update documentation</strong> if needed and I <strong>move tickets to Done</strong>.</p><p>The dynamic that emerged: I’m not trying to be a developer. I handle what I’m actually good at — requirements, testing, feedback, priorities. Claude handles implementation, debugging, technical decisions (with input from me when needed). It works because we’re not pretending the roles are interchangeable.</p><blockquote>It’s a real partnership. One partner just happens to be AI.</blockquote><h4>🎨 How I Design the App</h4><p>I don’t have Figma files. I start with a rough idea of what a feature should do and how it might look, then describe it to Claude. Sometimes I ask for suggestions, sometimes I have something specific in mind and describe the key parts. I never try to get everything perfect in one prompt — I start with the core, then add details one at a time.</p><p>Sometimes I borrow from other apps. The loading animation came from Shazam’s “listening” effect — I told Claude I wanted to replicate that but make it unique with emojis that appear and disappear based on the category you’re searching for.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o12D1c1U3JIkBH5aJZhGKA.jpeg" /><figcaption>Designing the app by pinpointing details from other apps I like — no Figma needed</figcaption></figure><p>When designing the compass, I downloaded a few compass apps and picked out small details I liked. I’d screenshot them and show Claude. Apple’s compass has a nice touch where the cardinal directions rotate to stay readable as the compass spins. I asked Claude to copy that.</p><p><strong>Having 10 years in UX probably makes this easier for me. But using other apps as references seems like a way anyone could get to something that looks decent.</strong></p><blockquote>I downloaded a few compass apps and picked out small details I liked. I’d screenshot them and show Claude.</blockquote><h4>🧰 What’s In My “Stack”</h4><ul><li><strong>Claude CLI in Warp Terminal</strong> — Claude can see your files, understand structure, read compilation errors. Warp’s AI answers command questions, which removes most of the Terminal anxiety.</li><li><strong>Xcode</strong>, with Claude’s guidance — I learned enough to build, test, and read errors. That’s about it.</li><li><strong>Git and GitHub</strong> as backup — commit often, push always. This saved me when dark mode broke everything, when the toilet categorization logic disappeared, and roughly 47 other times I’ve stopped counting.</li><li><strong>A kanban board</strong> (I use YouTrack) because when ideas pile up and bugs multiply, having structure helps. And a <strong>screenshot tool</strong> (I use CleanShot X) since you’ll be sharing a lot of visuals.</li></ul><p>What you don’t need: coding expertise, Terminal knowledge, years of experience. Just curiosity and enough discipline to test things before committing them.</p><h3>👍 What Principles Seem to Work &amp; Last</h3><p>AI tools will keep changing. Cursor, Lovable, Replit — all evolving quickly, probably merging or getting replaced by something better. But some things seem more durable:</p><ol><li><strong>Treat AI like something you’re working with, not something you’re commanding.</strong> It needs context and conversation, not perfect instructions.</li><li><strong>Plan before building.</strong> Look at what exists, propose something, poke holes in it, spot the risks, write it down, then build.</li><li><strong>Give it memory through files.</strong> CLAUDE.md and project documentation mean context carries across sessions. What you learn compounds instead of disappearing.</li><li><strong>Use version control</strong> so you can experiment without fear. Git means there’s always a working version to fall back to when things break.</li><li><strong>Keep some structure when things pile up.</strong> A kanban board or simple list creates focus when ideas and bugs start multiplying.</li><li><strong>The honest part: it’s still work.</strong> You need to think critically, give clear feedback, test thoroughly, learn enough about tools like Git and Xcode and Supabase (with help). You’ll wear different hats and put in hours.</li></ol><p>What’s different: you don’t need years of coding experience to go from an idea to something real. The barrier is lower than it looks.</p><blockquote><strong>One caveat: </strong>My UX background definitely helped a lot along the way. I did not have to design the app in Figma — I just described my vision to Claude; but it might be harder for someone with other background (like technical — which gives you different superpowers, though).</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JukbQ6sCPIl9NH4dBUnUXA.jpeg" /><figcaption>My kanban board — helps me stay organized, prioritize tasks, plan features ahead</figcaption></figure><p><strong>Credits:</strong> I have to mention three amazing people who helped me get on the right track:</p><ul><li><a href="https://www.linkedin.com/in/marek-ke%C5%99ka-24832b47/">Marek Kerka</a>, my colleague and AI evangelist who is always willing to provide helpful tips and expand horizons</li><li><a href="https://www.lennysnewsletter.com/podcast">Lenny Rachitsky</a> and his newsletter and podcast — you most likely know him for his great way of interviewing his guests, getting the most out of their knowledge for the good of the listeners. Many of his episodes inspired me and gave me the courage to try harder.</li><li><a href="https://www.youtube.com/@howiaipodcast">Claire Vo</a> and her podcast How I AI — you may also know her for her <a href="https://www.chatprd.ai/">ChatPRD</a> tool she made for fellow product managers. Every episode of her podcast is packed with practical examples and tips. It helped me tremendously.</li></ul><h3>🏃‍♂️ Building the App: The Last Mile</h3><p>Building the app was most of the work, but not all of it. The last stretch involved <strong>legal documents</strong> — Privacy Policy and Terms of Use for the App Store. Claude helped draft them considering GDPR and Apple’s requirements, then I made a simple website to host them.</p><p><strong>App Store Connect</strong> was new territory. Claude gave step-by-step instructions, and when I got confused I’d screenshot where I was and paste it in. Claude would walk me through the next steps.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MXSBTU4SqQKmsLvDaxiblQ.jpeg" /><figcaption>App metadata in App Store Connect, Landing page on my simple personal website, Privacy policy and more…</figcaption></figure><p><strong>The review cycles</strong> took 2–3 days of waiting each time. The reviewer had no issues with the code — just wanted Terms added, the onboarding flow adjusted for location permissions, some metadata fixes. I also put together a basic <a href="https://www.ondrejmachart.com/little-explorer"><strong>landing page</strong></a><strong> </strong>for the app.</p><p><strong>Then there was EU trader verification</strong> before I could release in Europe. Claude helped me understand what was needed and how to complete it. The app launched September 22nd for US and Asia. Global release happened end of September after verification cleared.</p><p>Not particularly complex work. Just a lot of waiting.</p><h3>🏗️ What Happens Next: Improving the App &amp; Taking On the New Challenge</h3><p>Little Explorer is live and growing slowly through App Store organic search. I’m watching how people use it and adjusting search quality through the remote config system. The first user was in the US, which immediately taught me the app only showed metric measurements. Americans prefer feet. Added imperial units in the next version.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6JMII1dyt5zd_CZKia2k4A.jpeg" /><figcaption>The app in the App store, my beloved “GottaGo mode” for Toilets finder and our story written in the app</figcaption></figure><blockquote>But this was never really about building a successful app. It was about seeing if I could build something complete — end to end, on my own, with AI helping.</blockquote><p>Now I’m ready to try again with that dividend investing app I failed at with Replit. It feels possible this time. Not because I learned to code, but because I learned how to work with something that can.</p><h3>💪 Your Turn: What Will You Build?</h3><p>You don’t need to start with your dream project. Start with something smaller — personally useful, complex enough to be a real test of whether this approach can work. Create a CLAUDE.md file from the beginning. Ask Claude to set up Git and some kind of task list. <strong>Have conversations instead of just firing off requests.</strong></p><blockquote>You’ll break things. Get stuck. That’s not failing. That’s just what it looks like.</blockquote><p><strong>Don’t sweat it. It doesn’t have to go viral overnight. </strong>Maybe you’re like me and you’re happy at your day job and just want to follow your builder’s heart without worrying about whether it will make any income at all. Or maybe you dare to dream big and want to give it try. Just remember that it’s rare to hit it right on the first try. That should not be the goal. The goal should be to do the work to learn the lessons that might get you there later.</p><p>Two months ago, I couldn’t code. Now there’s an app I built on the <a href="https://apps.apple.com/us/app/little-explorer/id6749822199">App Store</a>, used by parents in US, Hong Kong, Czechia, Germany, Netherlands and elsewhere.</p><p><strong>What will you build?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=42b45626f16e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>