<?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 Darren Lacy on Medium]]></title>
        <description><![CDATA[Stories by Darren Lacy on Medium]]></description>
        <link>https://medium.com/@darren.lacy?source=rss-241c4022c320------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*42PnSMJAQwHV6FJfn01GgQ.jpeg</url>
            <title>Stories by Darren Lacy on Medium</title>
            <link>https://medium.com/@darren.lacy?source=rss-241c4022c320------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 21:58:14 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@darren.lacy/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[Rewriting Brooks’ Law]]></title>
            <link>https://medium.com/@darren.lacy/rewriting-brooks-law-e0fef9790007?source=rss-241c4022c320------2</link>
            <guid isPermaLink="false">https://medium.com/p/e0fef9790007</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-coding]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[future-of-work]]></category>
            <dc:creator><![CDATA[Darren Lacy]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 15:01:03 GMT</pubDate>
            <atom:updated>2026-06-24T23:30:40.134Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vvJCNSGnh0lnGYz1Euxj0g.png" /></figure><p><strong>The Math Behind the Solo AI-Augmented Engineer</strong></p><p><em>Why one well-equipped engineer can credibly outperform a conventional pizza-sized team — and what the frontier labs’ own data tells us about the ceiling.</em></p><h3>A Note Before We Start</h3><p>What follows is a thought experiment about the theoretical upper bound of solo AI-augmented software development. It’s an edge case that requires nearly every condition to be aligned, and I want to be upfront: the numbers below are not what every engineer should expect on every project. They’re what becomes possible when the prerequisites stack up.</p><p>The reason I’m writing this now is that I happen to be running the experiment in real time. I’m a principal-level engineer with a career at a tech company building and supporting commercial products, including zero-to-one work at scale. The iOS app I’m about to launch on the App Store is greenfield, in a consumer productivity domain I know deeply. The stack is Swift/iOS — well-trodden territory the model has been thoroughly trained on. I’m working solo with an AI coding workflow I’ve been refining for months. Nearly every condition the model below depends on, I happen to have.</p><p>So I’m writing this in two voices: the analytical voice that lays out the framework and the citations, and the personal voice that admits I’m actually testing it. Whether I hit the upper end of the range is something the next few months will tell. Your mileage will vary — possibly significantly — based on how many of the prerequisites you actually hold. With that context set, here’s the framework.</p><h3>The Classical Constraint</h3><p>In 1975, Fred Brooks observed in <em>The Mythical Man-Month</em> that adding engineers to a late software project makes it later. The mechanism is communication overhead: the number of channels in a team of <em>n</em> members grows as <em>n(n−1)/2</em>. A team of eight carries 28 channels of potential coordination loss; a team of two carries one. Each channel is lossy — context degrades, intent gets misinterpreted, coordination consumes time, and every handoff is a place where something drops.</p><blockquote>Uncertainty is reduced through <strong><em>decisions</em></strong>, not through <strong><em>headcount</em></strong>.</blockquote><p>Two decades later, Steve McConnell quantified the downstream effect in <em>Software Estimation: Demystifying the Black Art</em> and <em>Software Project Survival Guide</em>. McConnell’s most cited contribution — the Cone of Uncertainty — captured the observation that estimation accuracy widens dramatically in early project phases and narrows only as scope and variability are removed through decisions. The deep implication, often missed, is that uncertainty is reduced through <em>decisions</em>, not through <em>headcount</em>. Larger teams pay a multiplicative coordination tax precisely when they need to move fastest through the wide part of the cone.</p><p>The practical consequence has been the same for fifty years: software organizations have grown teams to ship more, accepted the network cost as inevitable, and tried to compensate with process. Brooks’s Law was treated as a constraint to manage, not a constraint to break.</p><blockquote>AI does not occupy a node in the human communication graph — it expands the capacity of an existing node.</blockquote><h3>How AI Changes the Math</h3><p>A small team augmented with a well-orchestrated AI coding workflow attacks both problems simultaneously. By keeping the human node count low, the team retains the linear-cost coordination profile of a one- or two-person team while expanding effective execution capacity through AI-driven code generation, refactoring, and iteration. Critically, the AI does not occupy a node in the human communication graph — it expands the capacity of an existing node. This shifts the efficient frontier of the size-vs-capability tradeoff. The curve that previously forced teams to grow in order to ship more has moved.</p><p>The solo engineer-plus-AI configuration represents the theoretical limit case. With one human node, the communication graph collapses to zero inter-human channels — the <em>n(n−1)/2</em> coordination tax becomes literally zero rather than merely small. Decision velocity reaches its maximum possible value because there is no alignment cost, no consensus to build, no second taste to reconcile against. The Cone of Uncertainty narrows fastest in this configuration because every scope-defining decision is unilateral and instantaneous; the variability that McConnell identifies as the actual driver of estimation error is removed in real time as the engineer makes choices.</p><blockquote>This advantage is bought at a cost…</blockquote><p>This advantage is bought at a cost: a solo engineer retains only one unit of review-and-decide bandwidth — the very faculty that bottlenecks AI-augmented work at scale. The model below therefore assumes a best-in-class engineer: someone with strong product judgment, deep engineering experience, the taste to recognize when AI output is wrong or merely adequate, and the discipline to maintain rigor without peer review acting as a backstop. It further assumes a relatively known-scope greenfield project, a well-trodden technology stack with abundant training data, a product domain with cheap reversibility (consumer software more than regulated systems), and a mature AI orchestration workflow.</p><h3>The Multipliers</h3><p>Under these conditions, plausible multipliers for one best-in-class engineer plus AI, relative to that same engineer working alone without AI on the same project, fall into three regimes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hcWnhcfH0FT-sIR4PjSqFg.jpeg" /></figure><p>The <strong>conservative case</strong> reflects a defensible floor — even with friction, suboptimal context engineering, and meaningful review burden, a principal-level engineer with AI ships roughly three times faster than alone. The <strong>realistic case</strong> reflects the realized steady state for a mature solo workflow on appropriate work, consistent with both Anthropic’s internal data and broader top-quartile industry benchmarks. The <strong>peak case</strong> is realized when the engineer operates with extensive product surface area, a well-trodden stack, strong product taste reducing rework, and a workflow continuously optimized over the life of the project.</p><p>Agility — the ability to change direction, kill bad ideas, and ship iterations — scales further than raw throughput, because the elimination of the last human-to-human alignment cost is qualitatively significant rather than merely incremental. A solo engineer pivots at the speed of thought; even an aligned two-person team has a brief alignment beat on any direction change. This is the regime in which the solo AI-augmented engineer beats the conventional team most decisively — not on raw output, but on time-to-validated-direction.</p><h3>What the Frontier Labs Are Saying</h3><p>These multipliers are not theoretical. They sit within an emerging consensus across frontier AI labs and independent industry observation in late 2025 and early 2026.</p><p><strong>Dario Amodei, CEO of Anthropic</strong>, has publicly placed 70–80% odds on the first one-person billion-dollar company emerging in 2026, naming developer tools and proprietary trading among the most likely categories. <strong>Sam Altman</strong> has reported a standing betting pool among tech CEOs on the timing of this milestone.</p><p><strong>Anthropic’s published research</strong>, <em>How AI Is Transforming Work at Anthropic</em>, found that its engineers report 50% productivity gains using Claude in 60% of their work — a 2–3x increase in just one year — and that 27% of AI-assisted work consists of tasks that would not have been done otherwise. Anthropic’s engineering team grew 4x while individual engineer output increased by 200% over the same period, implying a roughly 8x compounded team-level productivity gain.</p><p><strong>Boris Cherny</strong>, who leads Claude Code at Anthropic and was previously one of Instagram’s most productive engineers, has not manually edited a line of code since November 2025 and ships 10 to 30 pull requests per day. <strong>Spotify</strong> has disclosed that its highest-performing engineers have stopped writing code manually. <strong>Anthropic shipped its Claude Cowork product in approximately ten days using Claude Code</strong>, and roughly 90% of Claude Code’s own codebase was written by Claude Code itself.</p><p><strong>Independent benchmarks</strong> from Larridin’s 2026 Developer Productivity report place healthy AI coding tool ROI at 2.5–3.5x average and 4–6x for top-quartile organizations, with an important caveat: AI coding tools produce real productivity gains, but they also produce the illusion of productivity gains, and traditional benchmarks cannot tell the difference. Disciplined measurement matters.</p><p>Solo founder case studies provide the upper-tail evidence. Pieter Levels operates multiple products at over $3M ARR with zero employees. Matthew Gallagher’s Medvi reached $401M revenue in its first year as a solo-founded company with an estimated $1.8B trajectory. Ben Broca’s Polsia crossed $1M ARR while managing 1,100 client companies solo.</p><h3>Honest Caveats</h3><p>Three apply specifically to the solo configuration.</p><p><strong>Sustainability. </strong>Solo work has a hard ceiling on hours that a team can extend through load-sharing. The multipliers above describe peak project velocity, not indefinite steady state. Burnout risk is real and the configuration does not eliminate it.</p><p><strong>Quality risk. </strong>With no human peer review, the engineer must rely on AI-mediated review and explicit personal discipline — test coverage, written design rationale, deliberate cooling-off before merging — to substitute for what a second engineer would catch organically.</p><blockquote>Human brains can hold a suprisingly large amount of data for immediate recall</blockquote><p><strong>Scope discipline. </strong>This configuration excels on bounded-scope work and degrades sharply as scope expands beyond what one head can hold in working memory. The model holds for a discrete product or major feature; it does not extrapolate to building a full platform organization. The engineer’s working memory becomes the binding constraint long before the AI’s capability does.</p><h3>The Absolute Peak: What’s Defensible at the Ceiling</h3><p>The numbers above describe the realistic operating envelope. There is a defensible ceiling that sits meaningfully higher, anchored to specific published frontier-lab data points rather than extrapolation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c2sx9dflHbz5QWcQUl9PUA.jpeg" /></figure><p><strong>12x throughput</strong> sits between two specific empirical anchors. First, Boris Cherny’s reported 10–30 PRs per day at Anthropic, which conservatively discounted for PR-equivalence still implies roughly 8–15x sustained throughput versus a high-performing engineer working without AI. Second, the disclosed fact that approximately 90% of Claude Code’s own codebase was written by Claude Code itself — meaning the human leverage factor on code volume alone is roughly 10x at one of the most demanding engineering organizations in the industry. Anthropic’s team-level 8x provides the floor; the solo configuration credibly exceeds this because it eliminates the residual coordination tax that even a high-trust team carries. Beyond 12x sustained throughput, claims cross from “consistent with reported data” into “exceeding what any frontier lab has publicly demonstrated as a steady state.”</p><p><strong>25x agility</strong> is anchored to Anthropic’s ten-day build of Claude Cowork. A comparable productivity-app launch from a conventional team would typically take four to six months, implying time-to-shipped-product compression of roughly 16–24x — and that is for a product that actually shipped to external users, not a prototype. For pure exploratory work — direction-finding, prototyping, killing bad ideas — the multiplier credibly extends to 25x because each discarded iteration costs near-zero in AI-augmented work, while in a conventional team each pivot triggers alignment meetings, refactor coordination, and sprint replanning. The solo-founder claims of “equivalent output to a 15–20 person team” are almost entirely an agility argument once decomposed, and those numbers come from real revenue-generating businesses, not theory.</p><p>Reaching the absolute peak requires every condition pegged at the maximum: a domain so well-trodden that the model has effectively been trained on adjacent problems; product surface area large enough for massively parallel exploration; workflow maturity measured in months of refinement; scope tight enough to fit in one engineer’s working memory; and cheap reversibility — early-stage product where you can ship, observe, and iterate within days.</p><blockquote>Brooks’s Law is not wrong, Brooks’s Law described a system in which adding capability required adding humans, and therefore required paying the coordination tax. That equation is now broken.</blockquote><h3>What This Means</h3><p>The honest takeaway is not that Brooks’s Law is wrong. It is that Brooks’s Law described a system in which adding capability required adding humans, and therefore required paying the coordination tax. That equation is now broken. Capability can be added without adding nodes to the human communication graph. The constraint has moved from team size to human review bandwidth.</p><p>For most consumer software work performed by experienced engineers with mature AI workflows, <strong>5x throughput and 10x agility versus working alone without AI</strong> is the defensible headline. <strong>12x and 25x</strong> define the ceiling — not theoretical, but anchored to what Anthropic, Spotify, and the broader frontier-lab ecosystem are publicly reporting in 2026.</p><blockquote>In an AI-powered world today, product leaders MUST ask the right question</blockquote><p>The implication for product leaders is that the right question is no longer “how big should the team be?” It is “what is the smallest team that can hold this scope in working memory, and how do we equip them?” For most product surfaces, that team is one or two people.</p><h3>Where I Stand in the Experiment</h3><p>I haven’t called success on this yet... I won’t until the app has been live in the App Store long enough for the market to weigh in. What I can say at this checkpoint is that my own measurements suggest I’m sitting somewhere between the typical upper bound (8x throughput, 15x agility) and the theoretical peak (12x and 25x). That’s promising. It is not yet validated.</p><p>Two things could prove me wrong, and I want to name both.</p><p>First, the project isn’t finished, its currently sitting at ~95% complete when I left for vacation last week. If my pace collapses in the last mile, the multipliers regress toward the middle of the range. That’s the throughput risk, and it’s the easier one to absorb.</p><blockquote>Speed without product-market fit is just failure on a shorter timeline</blockquote><p>Second, and more importantly: building fast is not the same as building well. If the product flops when it reaches users — if the design instincts I trusted turn out to be wrong, if the market doesn’t respond, if the differentiation isn’t real — then the agility multiplier becomes a measure of how efficiently I produced something nobody wanted. Speed without product-market fit is just failure on a shorter timeline. That said, reducing elapsed time between “fail-fast” cycles can have its own benefit.</p><p>So consider this article a checkpoint, not a conclusion. I plan to re-measure and write a follow-up in six months from now (or less), with the actual numbers — including the uncomfortable ones if that’s how it goes. If the app succeeds in the market and the multipliers hold, this framework picks up one data point of real-world validation from a working engineer outside a frontier lab. If the app fails, that’s also a data point, and worth publishing honestly.</p><p>Until then, the headline numbers stand on the published frontier-lab data, including my own actual, yet interim, throughput gains, and not YET on my actual final outcome. The theoretical case is sound regardless of how my specific experiment turns out. Whether the lived experience matches the theory — for me, and eventually for many others — is the question the next few months will answer.</p><p>Dario Amodei put 70–80% odds on the first one-person billion-dollar company emerging in 2026. I’m not under any illusion that I’ll be the one to claim that prediction — but I do want to get as close as I can to whatever ceiling actually exists for a solo engineer with the right setup. This might be my path to find out.</p><h3>References</h3><h3>Foundational literature</h3><p>• Brooks, F. P. (1975, rev. 1995). The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley.</p><p>• Boehm, B. W. (1981). Software Engineering Economics. Prentice-Hall.</p><p>• McConnell, S. (1997). Software Project Survival Guide. Microsoft Press.</p><p>• McConnell, S. (2006). Software Estimation: Demystifying the Black Art. Microsoft Press.</p><h3>Frontier lab statements and research</h3><p>• Amodei, D. (May 2025). Remarks at Anthropic Code with Claude developer conference, San Francisco. Reported by Inc. magazine. Reaffirmed publicly May 2026.</p><p>• Altman, S. Public remarks on the CEO betting pool for the first one-person unicorn.</p><p>• Anthropic. (2026). How AI Is Transforming Work at Anthropic. Research publication. <a href="https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic">https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic</a></p><p>• Anthropic. (2026). How Anthropic teams use Claude Code. Internal case studies, published PDF.</p><p>• Cherny, B. Lenny’s Podcast interview, February 2026. (100% of code via Claude since November 2025; 10–30 PRs/day; team grew 4x with 200% per-engineer output gain.)</p><h3>Industry data points</h3><p>• Spotify engineering disclosure, late 2025.</p><p>• Larridin. (2026). Developer Productivity Benchmarks 2026: AI-Native Engineering Data.</p><p>• Anthropic Claude Cowork launch reporting, January 2026.</p><h3>Solo founder case studies</h3><p>• Levels, P. — multiple products at $3M+ ARR, zero employees.</p><p>• Gallagher, M. / Medvi — $401M revenue year one, solo-founded; $1.8B trajectory as of May 2026.</p><p>Broca, B. / Polsia — $1M+ ARR managing 1,100 client companies solo.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e0fef9790007" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From “I’ll Just Build You One” to the App Store: What Shipping My First iOS App Taught Me About…]]></title>
            <link>https://medium.com/@darren.lacy/from-ill-just-build-you-one-to-the-app-store-what-shipping-my-first-ios-app-taught-me-about-96c779dca0e7?source=rss-241c4022c320------2</link>
            <guid isPermaLink="false">https://medium.com/p/96c779dca0e7</guid>
            <category><![CDATA[ios-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Darren Lacy]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 08:15:11 GMT</pubDate>
            <atom:updated>2026-11-09T17:33:58.315Z</atom:updated>
            <content:encoded><![CDATA[<h3>From “I’ll Just Build You One” to the App Store: What Creating My First iOS App Taught Me About Building in the Age of AI</h3><p>How a software engineer built his wife’s dream app in days — and what it revealed about one of the skills AI can’t currently replace.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1iCyDNLriUW2Xq31wSmZJg.png" /></figure><p>My wife has used the same to-do app for five or six years. It was the kind of app that just worked — quietly, reliably, invisibly. The kind of product you don’t think about because thinking about it would mean something had gone wrong.</p><p>Then she switched to iPhone.</p><p>The iOS version was missing one feature. Not a suite of features, not a broken experience — one. The feature she used every single day. The thing she loved most about the app. It existed on Android. It had never been added to iOS, and when I reached out to the PM who owned the product, the answer was honest if brutal: they were de-investing in the consumer space, pivoting toward enterprise. That feature wasn’t coming.</p><p>I looked at my wife and said, “I’ll just build you a new app.”</p><p>She laughed. I didn’t.</p><p>Three days later, she was using a build I’d shipped that afternoon. One of the first things I implemented was that feature — the missing one, the dealbreaker. It took me ten minutes.</p><p>I want to be careful about how I say this, because I’m not trying to take a shot at the team that built that app. Prioritization is real, and pivots are hard. But as an engineering leader, I know the difference between <em>can’t</em> and <em>won’t</em>. That feature wasn’t a technical problem. It was a choice. And it only takes one choice like that — one feature, one customer need quietly deprioritized — to lose someone who had been loyal for six years.</p><p>A few days after that first build, after I’d taken every piece of her feedback and shipped it — not in a sprint, not in a quarter, in <em>days</em> — she looked at me and said, “Oh my goodness, this is truly unbelievable.”</p><p>That moment is what this article is about.</p><p>Not because building iOS apps suddenly got easy. Not because the technical lift disappeared. But because something has fundamentally shifted in what’s possible when you combine genuine product intuition, real user feedback, and AI that can implement — <em>actually implement</em> — what you’re describing.</p><p>I’ve spent years leading software engineering teams building both internal and commercial products. I know what good looks like. I’ve also seen firsthand what happens when the people making product decisions are too far from the customer to feel the gap. What I didn’t know — what genuinely surprised me — is how much of the remaining barrier to shipping great software was just <em>implementation friction</em>. And that friction is collapsing.</p><p>This is the story of how I went from “I’ll just build you one” to having a product I’m proud to put my name on — and what it taught me about what it actually means to build software in 2026.</p><h3>Getting AI-Pilled</h3><p>A few months before I built this app, something shifted at work.</p><p>I joined a new team at Microsoft — one so deeply AI-native that it reframed everything I thought I knew about how software gets built. This wasn’t a team <em>talking</em> about AI. This was a team that had internalized it, shipped with it, and reorganized around what becomes possible when you stop treating AI as a tool and start treating it as a collaborator. That energy spread beyond my immediate team. I started identifying people — colleagues from current and prior teams, people outside Microsoft, some outside software engineering entirely — who were leaning into this shift in exciting ways. I proposed something simple: let’s meet bi-weekly, show each other what we’re learning, and build on each other’s work. The best learning communities are equal parts fun and purposeful — and there’s something deeply satisfying about that, especially at a company that believes the best work happens when you also build on the work of others.</p><p>What emerged was a much clearer picture of what people actually mean when they use terms like “agentic engineering” — and, on the flip side, “vibe-coded slop.” As a professional engineer, the distinction matters enormously to me. Vibe-coding produces exactly what the name implies. Agentic engineering — bringing genuine product vision, domain experience, and engineering judgment to bear while letting AI handle the implementation velocity — is something else entirely. It’s what I’d always wanted to do: build the products I could clearly see in my head, at a speed I never realized was possible.</p><p>One episode that crystallized this was Lenny Rachitsky’s conversation with Max Schoening, Head of Product at Notion: <a href="https://podcasts.apple.com/us/podcast/why-cultivating-agency-matters-more-than-cultivating/id1627920305?i=1000765889270">Why cultivating agency matters more than cultivating skills in the AI era</a>. The episode focuses largely on getting non-engineers to ship with AI — but as an engineer with a product mindset, I heard something different. I’m not limited by my ability to write code. I’m limited only by the clarity of my vision and the quality of my judgment. With AI, I can ship at a scale and speed that would have previously required an entire team. That’s not hyperbole. That’s the new reality.</p><p>But here’s the golden nugget I didn’t want to miss: just because AI can move faster than I ever could doesn’t mean I should let it outrun my understanding. From the very beginning, I instructed my AI coding assistant to bring me along — avoid jargon, explain the finer points, flag when the code isn’t adhering to SOLID design principles, and show me real examples so we could fix things together. Learning and shipping at the same time — faster than ever before — is a big part of what makes this so fun.</p><blockquote>AI carries the implementation velocity. I carry the judgment. And I refuse to outsource the understanding.</blockquote><p>The shift also crystallized through a handful of podcasts whose message, stripped down, was simple: <em>don’t be afraid. Do it. Learn, iterate, ship, be amazed.</em></p><p>Another episode that stuck with me — for very different reasons — was Lenny’s conversation with Caitlin Kalinowski, formerly of OpenAI, Meta, and Apple: <a href="https://podcasts.apple.com/us/podcast/why-were-at-the-beginning-of-the-ai-hardware/id1627920305?i=1000768249602">Why we’re at the beginning of the AI hardware boom</a>. Kalinowski returns repeatedly to themes that resonate deeply with me — truly understanding what the customer needs, what makes shipping a great product so hard, and why hiring people who are AI-focused in how they work is now non-negotiable. I spent time in hardware earlier in my career, and I’ve long admired what Apple has done fusing software and hardware into experiences that feel almost inevitable in hindsight. Great software requires exactly the same discipline. I’ll be exploring that connection — software architecture, engineering principles, and AI — in a future post. Consider this your teaser.</p><p>That cycle — learn, iterate, deliver, learn again — has brought something back that I hadn’t realized I’d been missing. I spend my days writing software at one of the largest technology companies in the world. Then I come home, and after time with my family, a workout, some time at the piano — I open my laptop and build more. Not because I have to. Because I want to. Because it’s that fun again.</p><p>I almost can’t stop. And I mean that as the highest possible compliment to what this moment in software development actually is.</p><h3>Design First (Even Without a Single Wireframe)</h3><p>Let me be honest about something: I didn’t open Figma. I didn’t write a spec. I didn’t sketch a single wireframe.</p><p>And I still designed first.</p><p>That’s the part I think most engineers miss when they hear “design-first mindset.” It doesn’t mean deliverables. It means discipline — the discipline to know exactly what you’re building, for whom, and why, before you write a single line of code. In my case, that meant three things: deep customer empathy, honest competitive research, and a relentless focus on what would make this product genuinely different.</p><p>My customer was sitting across from me at dinner every night. I already knew the product she’d been using for years — I used it too. So instead of a user research session, I had a real conversation. What did she always wish it did? What had she been doing manually for years because no app had ever bothered to solve it properly? What would make her actually prefer this over going back to something familiar?</p><p>That conversation gave me two things that became the soul of the product.</p><p>The first was the feature that dropped out of her old app when she switched to iPhone — a delight feature that made the experience feel intelligent and frictionless in a way the iOS version had simply abandoned. It took me ten minutes to build. I shipped it first, right after the core product was working, because I knew it was the reason she cared.</p><p>The second was something she’d been doing manually for years that I’d always found quietly remarkable — a deeply personal workflow she’d developed out of sheer necessity, because no app had ever thought to automate it. I built a feature that does exactly that. Once you set it up, it disappears into the background and just works — turning what used to be a tedious, error-prone manual process into something that feels almost like magic. The kind of feature that makes you wonder why nobody built it sooner.</p><p>Today when I took that app to the grocery store for the first time, I had one of those moments I won’t forget. I moved through the aisles efficiently, everything exactly where I expected it, the experience flowing in a way that felt almost choreographed. I was the most efficient I’d ever been on a grocery run — and I mean that literally, not as a figure of speech. Somewhere in the middle of that trip, I almost forgot I was the one who had built it. Almost.</p><p>That feeling — using your own product and being genuinely delighted by it — is something I’d heard described before but never experienced firsthand. It felt like what an author feels holding a printed copy of their book for the very first time. Proud, yes. But mostly just… amazed that it exists.</p><h3>The Part That Still Requires Being Human</h3><p>There’s a version of this story that gets told a lot right now, and it goes something like this: AI is democratizing software development. Anyone can build an app. The barrier to entry has never been lower.</p><p>That’s true. And it’s also, in an important way, beside the point.</p><p>Because the barrier that AI removes is the technical one. It doesn’t remove the harder one — the one that was always harder, even before AI existed. Understanding a human being well enough to build something they will genuinely love. Knowing what matters and what doesn’t. Having the taste to tell the difference between a feature that delights and one that just adds noise. Caring enough about the experience to get it right, even when getting it “good enough” would have been faster.</p><blockquote>AI can write code faster than any engineer alive. That’s not the achievement. The achievement is knowing what code to write — and just as importantly, knowing what to take away. Some of my most impactful sessions weren’t about adding features. They were about removing them, or refining something so subtle that most users would never consciously notice — but would absolutely feel.</blockquote><p>I also learned quickly that AI, left to its own devices, will happily invent solutions that violate platform conventions it absolutely should know. Apple has done extraordinary work codifying what makes great products great — Human Interface Guidelines, icon standards, interaction patterns that feel native because they <em>are</em> native. The moment I started explicitly instructing my AI to follow Apple’s best practices rather than improvise, everything got better. The results improved, and I stopped losing time to conversations that went something like: “Why didn’t you just do it the Apple way to begin with?” “Well, I didn’t realize I wasn’t until I wrote the code.” That particular exchange was equal parts maddening and illuminating. It taught me that being a great AI collaborator means being a great director — anticipating where it will drift and building guardrails before it does, not after.</p><blockquote>That’s the human stuff AI can’t replicate… yet! Judgment. Vision. Taste. Empathy for the person on the other end of the screen.</blockquote><p>I’ve spent my career building products for some of the largest companies in the world — some products were great, many just OK but useful — and the difference was almost never technical capability. It was whether someone in the room cared deeply enough about the customer to fight for the right experience, even when it was harder, even when the business pressures were pushing in another direction. None of that has changed. If anything, it matters more now. Because if AI lets anyone ship faster, the only remaining differentiator is judgment.</p><p>What I built for my wife isn’t remarkable because of how it was built. It’s remarkable — if it is — because it solves something real for a real person in a way that genuinely delights her. The technology is just what finally let me close the gap between the product I could imagine and the product I could ship.</p><p>That’s what I want every engineer reading this to take away. We’ve spent years wishing we could spend less time on implementation and more time on ideas, on customers, on the things that actually make products great. That wish just came true. The question is whether we’re going to use that freedom to build things that are genuinely worth building — or just ship slop faster than ever before.</p><p>I know which one I’m choosing.</p><p><em>In a future post, I’ll go deeper on what I learned about the iOS and watchOS ecosystem, the realities of publishing to the App Store, and how I thought about software architecture and engineering principles as the product matured. If the question of this post was “can someone like me build something worth using?” — the next one is “how do you make sure it’s built to last?”</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96c779dca0e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Glad to be on Medium]]></title>
            <link>https://medium.com/@darren.lacy/glad-to-be-on-medium-c9fee0c8a49b?source=rss-241c4022c320------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9fee0c8a49b</guid>
            <category><![CDATA[music]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[welcome]]></category>
            <dc:creator><![CDATA[Darren Lacy]]></dc:creator>
            <pubDate>Sun, 31 May 2026 20:09:45 GMT</pubDate>
            <atom:updated>2026-05-31T21:13:00.760Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*auoJ1gZlJMe80KBIfmUWjA.png" /></figure><p>Hey everyone. This is my first post on Medium. Really excited to be here as a new member. I have followed friends and colleagues who have published here for years, and I look forward to documenting my journey as a newly “AI-pilled” engineer working on a lot of projects spanning commercial, open source and hobby.</p><p>I’ll probably also talk a little bit about my experience as an engineer at Microsoft, AI topics in general, and whatever else strikes my fancy. As a professionally trained and educated musician/musicologist, I’ll also talk about the work I am doing in music, which is not totally independent from the software world I now live in.</p><p>Please feel free to reach out to me on LinkedIn or GitHub. I hope to earn some followers and add to my network.</p><p><em>Darren</em></p><p>LinkedIn: <a href="https://www.linkedin.com/in/darrenlacy/">https://www.linkedin.com/in/darrenlacy/</a><br>Github: <a href="https://github.com/Opus110ai">https://github.com/Opus110ai</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9fee0c8a49b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>