<?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 Leroy Leow on Medium]]></title>
        <description><![CDATA[Stories by Leroy Leow on Medium]]></description>
        <link>https://medium.com/@leroyleowdev?source=rss-85694527171------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*OHNGx1Y9rsD4-xHbwXLOVQ.jpeg</url>
            <title>Stories by Leroy Leow on Medium</title>
            <link>https://medium.com/@leroyleowdev?source=rss-85694527171------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:39:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@leroyleowdev/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[Stop Claude from Shipping Generic AI Design Slop (With Taste Skill)]]></title>
            <link>https://medium.com/@leroyleowdev/stop-claude-from-shipping-generic-ai-design-slop-with-taste-skill-de7b7fda916e?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/de7b7fda916e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-slop]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 13:16:01 GMT</pubDate>
            <atom:updated>2026-10-12T13:16:01.484Z</atom:updated>
            <content:encoded><![CDATA[<p>We’ve all seen it.</p><p>You ask Claude (or Claude Code) to design a landing page, dashboard, or portfolio site… and what comes back looks like every other AI-generated interface: centered hero, three identical feature cards, purple gradient, Inter font, slate-900 text, and zero personality.</p><p>It’s not that Claude is incapable. It’s that its default training biases it toward safe, average, text-first outputs. Visual taste — spacing rhythm, hierarchy, intentional asymmetry, restrained motion — is under-trained compared to pure reasoning and code generation.</p><p>The good news? There’s a free, open-source fix that’s become one of the most popular skills in the Claude Code ecosystem: <strong>Taste Skill</strong>.</p><h3>What Is Taste Skill?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*l3QR8s4Zh2Xlh0Cu" /><figcaption>Photo by <a href="https://unsplash.com/@villxsmil?utm_source=medium&amp;utm_medium=referral">Luis Villasmil</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Taste Skill (from the <a href="https://github.com/Leonxlnx/taste-skill">Leonxlnx/taste-skill</a> repository) is a set of carefully written SKILL.md instruction files that give Claude actual design judgment.</p><p>Instead of defaulting to the most common patterns in its training data, the skill forces Claude to:</p><ul><li>Read the brief first and infer a deliberate design direction</li><li>Choose a coherent visual system</li><li>Avoid the classic AI clichés (purple gradients, three equal cards, em-dashes everywhere, fake screenshots made of divs, etc.)</li><li>Run a pre-flight review before shipping code</li></ul><p>The primary skill is called design-taste-frontend. There are also specialized variants for minimalist, brutalist, soft, redesign, and image-to-code workflows.</p><h3>How to Install It</h3><p>The fastest way (works with Claude Code, and often Cursor/Codex too)</p><pre>npx skills add https://github.com/Leonxlnx/taste-skill --skill &quot;design-taste-frontend&quot;</pre><p>Or install the whole suite:</p><pre>npx skills add https://github.com/Leonxlnx/taste-skill</pre><p>After installing, restart Claude Code or start a new session so it picks up the skill.</p><h3>How to Use It</h3><p>Once installed, you can trigger it in two ways:</p><ol><li>Explicitly with the slash command:<br>/design-taste-frontend</li><li>Just describe a design task — the skill often activates automatically when the request matches its description.</li></ol><h4>The Three Dials (Your Secret Weapon)</h4><p>Taste Skill gives you three numeric controls (1–10) that act like equalizers for the design:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/891/1*RBvmkSTuMtbF5WF0Y0NdyQ.png" /></figure><p>You can also just say things like “make it more minimalist” or “increase the visual density” and the skill will adjust accordingly.</p><p>For redesigns, the skill is instructed to <strong>audit first</strong> before making changes — highly recommended.</p><h3>How to Check If Taste Skill Is Actually Installed</h3><p>Don’t guess. Verify:</p><p><strong>Inside Claude Code:</strong></p><ul><li>Type / and look for design-taste-frontend in the autocomplete</li><li>Type /skills or ask: “What skills do you have available?”</li><li>Try invoking /design-taste-frontend directly</li></ul><p><strong>On disk:</strong></p><pre>ls ~/.claude/skills/          # global/personal skills<br>ls .claude/skills/            # project-level skills</pre><p>Or use the skills CLI:</p><pre>npx skills list<br># or<br>npx skills ls -a claude-code</pre><p>If the folder exists but Claude doesn’t see it, make sure SKILL.md sits directly inside the skill folder (not nested one level deeper) and restart the session (or run /reload-skills).</p><h3>Quick Tips for Better Results</h3><ul><li>Give a clear brief: audience, product context, brand constraints, and references help a lot.</li><li>Start with one page or section rather than an entire site.</li><li>Ask for a one-line “Design Read” and a before/after comparison when redesigning.</li><li>Iterate by naming specific dimensions (“lower the density,” “strengthen the hierarchy,” “reduce motion”).</li><li>Pair it with Anthropic’s official frontend-design skill or image-to-code variants for even stronger results.</li></ul><h3>Final Thought</h3><p>Claude is already excellent at structured reasoning and clean code. Taste Skill simply teaches it to stop shipping the design equivalent of Lorem Ipsum.</p><p>Install it once, and you’ll notice the difference immediately: interfaces that feel intentional instead of generated.</p><p>If you’ve been frustrated by generic AI UI, this is one of the highest-leverage 30-second upgrades you can make right now.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de7b7fda916e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Agent Doesn’t Need to Be Smarter. It Needs to Be the Right Level.]]></title>
            <link>https://medium.com/@leroyleowdev/your-ai-agent-doesnt-need-to-be-smarter-it-needs-to-be-the-right-level-581bd7283070?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/581bd7283070</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[framework]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 07:51:01 GMT</pubDate>
            <atom:updated>2026-10-11T07:51:01.534Z</atom:updated>
            <content:encoded><![CDATA[<h3>Your AI Agent Doesn’t Need to Be Smarter. It Needs to Be the Right Level.</h3><p>Oct 9, 2026 · @Create a Web base calculator</p><p><em>A plain-English guide to the five levels of AI agents, adapted from the Agentic AI Progression Framework by Bornet et al., and why the simplest agent usually wins.</em></p><h3>Every vendor says “agent” now</h3><p>Open any SaaS pricing page and you’ll find an “AI agent”. Some are genuinely impressive. Some are a macro with a chatbot taped to the front.</p><p>The word has stopped meaning anything, which is a problem when you’re the one signing the contract or building the thing.</p><p>Here’s the distinction that matters. <strong>Traditional automation is a vending machine</strong>: press B4, get crisps, every time. The moment something unexpected happens, it fails. <strong>Agentic AI is closer to a self-driving car</strong>: it perceives what’s around it, plans a route, acts, adjusts to the road and learns from each trip.</p><p>Both are useful. They are not the same thing, and most products sit somewhere in between.</p><p>I’ve been working through two frameworks from Bornet et al. that cut through the noise: <strong>SPAR</strong>, which explains what an agent does, and the <strong>Agentic AI Progression Framework</strong>, which tells you how well it does it. Together they’re the most useful vocabulary I’ve found for talking about agents with both engineers and executives.</p><h3>SPAR: what every agent actually does</h3><p>Every real agent runs the same four-step loop: <strong>Sense, Plan, Act, Reflect</strong>. The name is a nod to a sparring partner, constantly reading and adapting to the other side.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*8l8x7VgYgqzCno6LQe0VKA.png" /></figure><ul><li><strong>Sense</strong> is the eyes and ears. The breadth and quality of its inputs cap how well it understands the situation.</li><li><strong>Plan</strong> is where agents differ most. Simple ones follow rules; advanced ones reason through several steps.</li><li><strong>Act</strong> separates agents from analytics dashboards. Good agents watch their own actions and correct course, and there’s always a way for a human to review and intervene.</li><li><strong>Reflect</strong> is the hardest part: folding lessons back into what the agent knows so it does better next time.</li></ul><p>The value is in the loop, not any single step. Sensing feeds planning, planning guides action, action produces experience, and reflection improves the next round. That loop is the line between an agent and a script.</p><p>When I evaluate an agent product now, I go stage by stage. What can it see? How does it decide? What tools can it actually use? And does it really learn, or just remember?</p><h3>The six levels, borrowed from self-driving cars</h3><p>SPAR tells you a car needs steering and brakes. It doesn’t tell you whether you’re in a 2005 sedan or a robotaxi. So Bornet et al. built a level system that deliberately mirrors the SAE’s six levels of driving automation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*APHaaSVP1ZDqKJWkcaD-6w.png" /></figure><p>The most practical change as you climb is <strong>how you talk to the agent</strong>:</p><ul><li><strong>Level 1–2:</strong> “Open email, click New Message, paste this address…”</li><li><strong>Level 3:</strong> “Write and send an email with this information, adapted to the recipient.”</li><li><strong>Level 4:</strong> “Handle all client communications so they stay well informed.”</li></ul><p>You move from writing procedures to setting objectives. That’s a management shift, not just a technical one.</p><p>And a reality check: as of Bornet et al.’s March 2025 reference, <strong>most agents on the market are Level 2 or 3</strong>, with a few specialised systems reaching Level 4 in narrow domains. Level 5 needs breakthroughs we haven’t had. If a vendor pitches you Level 5, you’ve learned something about the vendor.</p><h3>It’s a menu, not a ladder</h3><p>This is the part most people get wrong, and the framework is explicit about it: <strong>higher is not better.</strong> The levels are a catalogue of agent types, each suited to different jobs.</p><p>Plenty of drivers prefer plain cruise control even when highway autopilot is available. It’s predictable. The same goes for business: many tasks are best served by a Level 1 or 2 agent.</p><p>Hence the framework’s <strong>Golden Rule: the simpler, the better.</strong> Pick the simplest level that does the job adequately. It isn’t minimalism for its own sake; it’s about balancing capability against control.</p><p>I think of it as staffing a kitchen. You don’t hire a head chef to cook rice. A rice cooker is cheaper, faster and does the same thing every single time. A head chef earns their keep running a whole menu, adapting to what’s in stock and what diners send back.</p><p>Matching tasks to levels avoids two expensive mistakes: <strong>over-engineering</strong> (a Level 3 agent where a Level 1 script would do) and <strong>under-specifying</strong> (expecting a Level 1 bot to handle Level 3 work).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4rY5fC5kH_hBE2fmYQ7pEw.png" /></figure><p>As a developer, this is the most liberating idea in the framework. Not every problem needs an LLM. Sometimes the best agent is a well-written script.</p><h3>Autonomy costs you control</h3><p>Every step up the levels trades direct human control for autonomy. It’s like delegating to a new hire: step-by-step instructions at first, broad goals once you trust them. The difference is that a machine’s mistakes can be faster and harder to undo.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2OdbUH0Q9ODGrN0zY_wobA.png" /></figure><p>So for anything at Level 3 or above, the framework says to answer five questions <strong>before</strong> you scale:</p><ol><li><strong>Accountability.</strong> Who’s responsible when the agent gets it wrong? Decide before the incident, not after.</li><li><strong>Bias and fairness.</strong> Are the data and decisions equitable?</li><li><strong>Transparency.</strong> Can a human explain why the agent did what it did?</li><li><strong>Control systems.</strong> Is there an off-switch, an audit log and an escalation path?</li><li><strong>Compliance.</strong> Does it fit the rules for your industry and jurisdiction?</li></ol><p>Level 4 agents should also be <strong>domain-limited</strong>, the agent version of geo-fencing a self-driving car to mapped roads.</p><p>The framework’s framing is right: governance isn’t bureaucracy. It’s the infrastructure that lets you trust an agent enough to give it more autonomy. Without it, more autonomy is just more risk.</p><h3>Where I’d push back</h3><p>The framework is the clearest one I’ve seen for agent capability. Some caveats from someone who builds these things:</p><ul><li><strong>Ask vendors how their agent “reflects”.</strong> The framework calls Reflect the most sophisticated capability and credits Level 3 with reinforcement learning. Many real LLM-based agents do little true learning at runtime; “memory” often means a stored conversation log. Make them show you.</li><li><strong>Real products mix levels.</strong> A common pattern is an LLM front end (Level 3) deciding what to do, with an RPA bot (Level 1) doing the clicking. Classify what the <em>task</em> needs, not what the marketing says.</li><li><strong>Some examples sit uneasily with regulated work.</strong> Medical diagnosis and financial advice agents appear as Level 4 examples. In regulated settings, I’d design those to <em>recommend</em> while a qualified human decides.</li><li><strong>“Faster to deploy” is half the story.</strong> Higher levels go live faster technically, but the governance they need takes real time. Total time-to-value may not be shorter.</li><li><strong>The market snapshot will date.</strong> “Mostly Level 2–3” reflects March 2025. The framework will outlast that number.</li></ul><h3>Start simple, scale smart</h3><p>The framework doubles as a roadmap. Treat it as organisational change, not a tech rollout:</p><ol><li><strong>Educate leadership.</strong> Use SPAR and the levels as a shared vocabulary, so “agent” means the same thing in every meeting.</li><li><strong>Find low-risk, high-impact processes.</strong> Repetitive work with structured data is the natural starting point.</li><li><strong>Start at Level 1 or 2.</strong> Predictable behaviour and high human control build familiarity and trust.</li><li><strong>Set up governance.</strong> Form a cross-functional team (legal, compliance, IT, ethics) with policies for data access, decision authority and escalation.</li><li><strong>Pilot Level 3+ in a sandbox.</strong> Measure outcomes, tighten governance, and scale only what’s proven.</li></ol><h3>The bottom line</h3><p>The question isn’t “how smart is your agent?” It’s “is this the right level for this job, and can we govern it?”</p><p>Most of the time the answer is simpler than the demo suggests. A rice cooker beats a head chef at making rice.</p><p>Start small. Think big. Scale wisely.</p><p><em>What level are the agents you’re building or buying today? And has anyone tried to sell you a Level 5? I’d love to hear about it in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=581bd7283070" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Great Products Don’t Disrupt Industries. Systems Do.]]></title>
            <link>https://medium.com/@leroyleowdev/great-products-dont-disrupt-industries-systems-do-227ac67eb30b?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/227ac67eb30b</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[lean-canvas]]></category>
            <category><![CDATA[business-analysis]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Sat, 10 Oct 2026 16:51:01 GMT</pubDate>
            <atom:updated>2026-10-10T16:51:01.439Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Lean Canvas, first principles, the Ten Types of Innovation and platform thinking: four tools that work far better together than apart.</em></p><h3>The best product rarely wins on its own</h3><p>Every developer has watched a technically better product lose to a worse one. Better code, better UX, better performance, and still nobody switched.</p><p>It took me a while to accept why. Disruption almost never comes from the product alone. It comes from a <em>system</em>: the right problem, framed from scratch, solved across the whole business, and scaled through a model that gets stronger as it grows.</p><p>I’ve been working through a <strong>Disruptive Product Innovation Framework</strong> that packages exactly that system into four pillars:</p><ul><li><strong>Lean Canvas</strong> to find and frame the problem</li><li><strong>First Principles Thinking</strong> to strip it down to what’s actually true</li><li><strong>The Ten Types of Innovation</strong> to solve it across the whole business, not just the product</li><li><strong>The Platform Business Model</strong> to scale it</li></ul><p>None of these is new on its own. What’s useful is how they connect.</p><h3>Four pillars, one loop</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*PNv4R5jdH95Kv_FiBvtc6g.png" /></figure><p>The Lean Canvas surfaces the problem and your riskiest assumptions. First principles breaks that problem into fundamentals. Those insights feed the Ten Types of Innovation, so the solution goes beyond a feature list. The platform model then gives it a structure to scale.</p><p>And it doesn’t stop there. What you learn at scale feeds back into the canvas, and the loop runs again.</p><h3>Pillar 1: Lean Canvas, the business plan that fits on one page</h3><p>Forget the 40-page business plan. A Lean Canvas captures your idea on a single page so you can test it fast and pivot cheaply. Everything revolves around one thing: a <strong>unique value proposition (UVP)</strong> grounded in a real customer problem.</p><p>This framework’s version has nine boxes, filled in this order:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/991/1*dfkxyPbGBDhUkQjoi2r4rA.png" /></figure><p>Two things I like here. First, the problem comes before the solution, which is the opposite of how most of us developers start. Second, the <strong>unfair advantage</strong> box forces an honest conversation about your moat. “We’ll execute better” doesn’t count.</p><p>One note if you know the original Lean Canvas by Ash Maurya: it has a <strong>Key Metrics</strong> box and no partner box. This version swaps them, reflecting its platform focus. I’d still track your key metrics somewhere.</p><h3>Pillar 2: First principles, or stop copying the industry</h3><p>Most innovation is reasoning by analogy: look at what exists and make it 10% better. First principles thinking does the opposite. Break the problem down to what you know is fundamentally true, then rebuild a solution from there, ignoring how things are usually done.</p><p>The framework pairs it with the Lean Canvas in six moves:</p><ol><li><strong>Define the problem</strong> from its fundamentals, not from existing solutions.</li><li><strong>Ideate solutions</strong> on the canvas built on those truths, not on assumptions.</li><li><strong>Shape a UVP</strong> rooted so deeply in the core problem that it’s hard to copy.</li><li><strong>Test assumptions</strong> with customers, making sure each one is well reasoned.</li><li><strong>Iterate or pivot.</strong> If feedback undermines a fundamental, be willing to rethink radically.</li><li><strong>Plan for scale</strong> by identifying what’s essential to growth.</li></ol><p>The classic example is <strong>Tesla</strong>. Rather than asking how to make a slightly better electric car, it broke the problem down to energy density, battery cost and materials, and attacked range, cost and performance from there.</p><p>It’s a good example, with a caveat: Tesla also had capital, timing, brand and policy tailwinds. First principles was necessary, not sufficient.</p><p>For developers, the everyday version is asking “what is this feature actually <em>for</em>?” before copying a competitor’s implementation.</p><h3>Pillar 3: The Ten Types of Innovation, or the product is only 2 of 10</h3><p>This is the pillar that changed how I think about my own projects. The Ten Types of Innovation sorts innovation into three groups, and <strong>only two of the ten types are about the product itself.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YT2BVfNwcjYdfohM0uCdfA.png" /></figure><p>That’s <strong>4 + 2 + 4</strong>. If your whole roadmap lives in rows 5 and 6, you’re competing on the easiest dimension to copy.</p><p>The real power comes from <strong>combining types</strong>. A rival can clone a feature in a sprint. Cloning a new profit model plus a partner network plus a community is much harder. First principles helps you pick which combination fits your problem.</p><h3>Pillar 4: Platforms, the multiplier</h3><p>A traditional <strong>linear business</strong> is a pipeline: make it, sell it, ship it. It’s rigid and expensive to scale. A <strong>platform</strong> doesn’t make the thing. It connects the people who make it with the people who want it, in a <strong>two-sided market</strong>, and earns from the exchange.</p><p>The easiest way I’ve found to explain this is a Singapore <strong>hawker centre</strong>.</p><ul><li><strong>Management doesn’t cook.</strong> It owns no food and no inventory. It connects stallholders with diners. That’s the difference between a platform and e-commerce: a restaurant cooks and sells, a hawker centre hosts.</li><li><strong>Network effects drive it.</strong> More good stalls attract more diners, and more diners make the rent worth paying, which attracts more stalls. That’s an <em>indirect</em> network effect, one side attracting the other, exactly like Uber’s drivers and riders. A lively centre also draws people simply because others are there, a <em>direct</em> effect.</li><li><strong>Critical mass is the hard part.</strong> On opening day, nobody comes because there are only five stalls, and no stalls join because nobody comes. You have to subsidise the start: rent discounts and promotions. Facebook and Booking.com spent heavily on early acquisition for the same reason.</li><li><strong>Once it’s full, it’s sticky.</strong> Regulars have “their” chicken rice stall and their usual table. Those are <strong>switching costs</strong>.</li><li><strong>Costs are externalised.</strong> Stallholders buy their own woks and perfect their own recipes, the way YouTube creators make the content YouTube serves.</li></ul><p>On top of that, platforms run on <strong>data and AI</strong>. Netflix and Spotify personalise everything to deepen engagement, and dynamic pricing balances supply and demand in real time. Revenue comes from many directions: transaction fees, subscriptions, advertising, premium services, data licensing.</p><p>This is why the framework treats platforms as the culmination. A platform amplifies every other innovation you’ve built.</p><h3>Six ways platforms fail</h3><ol><li><strong>Confusing a platform with e-commerce.</strong> If you own the inventory and sell it, you’re a shop.</li><li><strong>Never reaching critical mass.</strong> Without enough users on both sides, network effects never start.</li><li><strong>Ignoring governance and trust.</strong> One bad experience can empty a marketplace. Clear rules, enforced consistently, are the product.</li><li><strong>Letting network effects run wild.</strong> Uncontrolled growth means overcrowding and falling quality.</li><li><strong>Standing still.</strong> Static platforms become obsolete.</li><li><strong>Leaving monetisation too late.</strong> Diverse revenue from the start funds the innovation and retention you’ll need.</li></ol><h3>Where I’d push back</h3><ul><li><strong>Network effects aren’t profits.</strong> Uber took many years to make money. And when users happily run several apps side by side, as riders and drivers do across ride-hailing apps, switching costs collapse and the moat thins.</li><li><strong>Not everything should be a platform.</strong> It only fits when there are two sides with something to exchange. Plenty of great businesses are linear by design.</li><li><strong>The examples blur the lines.</strong> Amazon is cited as a platform, but it’s a hybrid: a retailer that owns inventory <em>and</em> a marketplace.</li><li><strong>Externalised costs come with externalised risk.</strong> User-generated content brings moderation and liability problems. That’s another reason governance matters.</li><li><strong>Lean Canvas is built for startups.</strong> Inside a large company you also have to deal with existing customers, brands and cannibalisation.</li></ul><h3>The bottom line</h3><p>A great product is one of ten ways to innovate. Disruption comes from the system around it: the right problem, rebuilt from first principles, innovated across the whole business, and scaled through a model that grows stronger with every user.</p><p>Which pillar is your current project weakest on?</p><p><em>If you’re building a platform, I’d love to hear how you’re tackling the chicken-and-egg problem. Tell me in the comments.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=227ac67eb30b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Asking What AI Agents Can Do. Ask Where They Should.]]></title>
            <link>https://medium.com/@leroyleowdev/stop-asking-what-ai-agents-can-do-ask-where-they-should-85b403384cf5?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/85b403384cf5</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[framework]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Sat, 10 Oct 2026 07:41:01 GMT</pubDate>
            <atom:updated>2026-10-10T07:41:01.405Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A practical, scored method for picking agentic AI projects that actually pay off, adapted from the Three Circles framework by Bornet et al.</em></p><h3>Everyone wants agents. Few pick the right job for them.</h3><p>AI agents are the most exciting thing I’ve built with in years. Software that can observe, plan and act on its own to finish a multi-step task is genuinely new.</p><p>But I’ve seen the same movie with every wave of enterprise tech. Companies rush in and launch pilots everywhere. Six months later they have <strong>pilot fatigue</strong>: budgets gone, nothing in production, and a leadership team that now thinks “AI doesn’t work here.”</p><p>The problem usually isn’t the technology. It’s the <em>selection</em>.</p><p>I’ve been studying a structured method for this, built on the <strong>Three Circles of Agentic Opportunity</strong> from Bornet et al. Its core idea is simple and slightly counterintuitive: <strong>the goal isn’t to automate as much as possible. It’s to automate what matters most.</strong></p><p>Here’s how it works, and where I think it falls short.</p><h3>Start with where agents don’t belong</h3><p>The framework opens with an unusual move: before hunting for opportunities, rule out the bad fits. An early, visible failure can poison adoption across the whole company, so this step deserves as much care as the search itself.</p><p>Five kinds of task are consistently poor fits for agents:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1020/1*4OC4oLPot37SmSQ-KDAeDQ.png" /></figure><p>My favourite part is the litmus test. <strong>Can you write the task as a standard operating procedure without saying “use your judgment”?</strong> If yes, it’s a candidate. If every second line says “it depends”, leave it with a human, at least for now.</p><p>As a developer, I’d add: if you can’t write the SOP, you definitely can’t write the prompt.</p><h3>The Three Circles: matter, possible, worth it</h3><p>Every candidate task gets tested against three questions. The sweet spot is where all three overlap.</p><p><strong>1. Impact: “Will it matter?”</strong> This isn’t just hours saved. It’s what those hours unlock. Ask: if this were automated tomorrow, what could the team do instead? The best targets are often boring: data entry, report compilation and routine monitoring that keep skilled people from strategy, problem-solving and client work.</p><p><strong>2. Feasibility: “Can it be done?”</strong> Think of it as checking you have the ingredients before you cook. Feasible tasks have clear, repeatable rules; structured data behind modern APIs; success criteria you can check automatically; manageable consequences if something goes wrong; and room for a human to review before final action.</p><p><strong>3. Effort: “Is it worth it?”</strong> This covers the practical realities, technical and organisational. Is the process documented? Is the team open to change? Can you start with a small pilot? Do the benefits justify the cost? Can core operations keep running during rollout?</p><p>Miss a circle and you get a predictable failure:</p><ul><li>High impact, low feasibility is <strong>aspirational</strong>, not actionable.</li><li>Easy but low value is a <strong>distraction</strong>.</li><li>Only the overlap is a project worth starting.</li></ul><p>The framework’s guidance for a first project is spot on: ambitious enough to show real value, but not so complex that failure kills confidence.</p><h3>Turning circles into numbers: the 35-point score</h3><p>Circles are great on a whiteboard. To actually decide, the method scores every task on a 1–5 scale across seven criteria:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zX9pbf1UUBdWY-aiZ7aa8A.png" /></figure><p>Two details matter:</p><ul><li><strong>Effort is reverse-scored</strong>, so every number points the same way. Higher always means better.</li><li><strong>An Impact score of 15 or more</strong> marks a strong candidate.</li></ul><p>The scoring starts with a two-hour team workshop to build a task inventory. Skip open brainstorming and ask pointed questions instead: What do you repeat every week? What keeps you from strategic work? Where are the bottlenecks? Which routine tasks need the most checking? Log how often each task happens and how long it takes.</p><p><strong>A worked example.</strong> A marketing agency (“Jenny’s agency” in the source material) spent <strong>over 200 hours a month</strong> on client reports. Worse, that work kept analysts from giving clients the strategic insight they actually wanted. Scored, it looked like this:</p><ul><li>Impact <strong>19/20</strong>: huge time sink, blocked strategic work, error-prone, highly repeatable</li><li>Feasibility <strong>8/10</strong>: well documented, data reachable through APIs</li><li>Effort <strong>4/5</strong>: standard tools, minimal customisation</li><li>Total: <strong>31/35</strong></li></ul><h3>Seeing it: the prioritisation matrix</h3><p>Scores are rigorous but hard to present to a leadership team. So the method plots every task on a 2×2: business impact across, complexity of transformation up.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*StQkb23ulLuE9hZkSVnN4Q.png" /></figure><p>Quick Wins (high impact, low complexity) are your first pilots. Strategic Projects (high impact, high complexity) get planned for later phases. Low Priority tasks are done when convenient. Avoid is exactly what it says, however cool the tech.</p><p>Jenny’s client reports landed squarely in Quick Wins and became the agency’s first agent. Its success built the momentum for <strong>five more agent implementations</strong> over the next year. That’s the real return on a good first pick: it buys you the trust to do the next one.</p><h3>Four traps I recognise from real projects</h3><p><strong>Automating roles instead of tasks.</strong> A job is a bundle of connected tasks full of context-switching, judgment and collaboration. Agents can’t replace that bundle. Ask “Can this specific step be agentic?”, not “Can we replace this job?”</p><p><strong>Chasing visible over valuable.</strong> Leaders tend to pick the tasks they see most often. Those are rarely the ones that matter most. Scoring exists to beat that bias.</p><p><strong>Skipping documentation.</strong> If a process isn’t written down, an agent will be unreliable at it. Good SOPs cut build time and raise success rates. Treat documentation as step zero, not an afterthought.</p><p><strong>Trusting your first idea.</strong> Companies arrive knowing which agent they want. In nearly half of cases, it’s not the most valuable one. One client was set on automating social media posting; scoring showed lead qualification would deliver <strong>five times the ROI</strong>.</p><h3>Where I’d push back</h3><p>The method is solid. A few things are worth knowing before you adopt it as gospel.</p><ul><li><strong>Impact dominates the score.</strong> It’s 20 of 35 points (57%). Effort is only 5 (14%). A technically brutal project can still score well if it’s valuable enough.</li><li><strong>The scores measure less than the circles describe.</strong> The Feasibility circle lists five traits, but only two get scored. The Effort circle talks about team readiness, but the Effort score only rates technical complexity. Error consequences, human review and change appetite are left to judgment, so discuss them explicitly.</li><li><strong>There’s no pass mark for the total.</strong> Impact has a threshold (15+). The composite doesn’t. Decide yours before you score, not after.</li><li><strong>The evidence is anecdotal.</strong> “Nearly half” and “5× ROI” are field stories, not controlled studies. Useful, but not proof.</li><li><strong>Mind the y-axis.</strong> It’s labelled complexity but built from scores where high means easy. It’s easy to plot things upside down.</li></ul><h3>The playbook</h3><p>Here’s the whole method as something you can run with your team next week:</p><ol><li><strong>Screen out the red flags.</strong> Drop anything that needs empathy, judgment under ambiguity, legal authority or messy collaboration. Apply the SOP test.</li><li><strong>Run a task inventory.</strong> Two hours, targeted questions, and a log of frequency and time per task.</li><li><strong>Score every task.</strong> Impact out of 20, Feasibility out of 10, Effort out of 5. Add them up for a score out of 35.</li><li><strong>Plot them on the matrix.</strong> Find your Quick Wins.</li><li><strong>Pick one or two pilots.</strong> Beyond the score, choose ones that can show measurable results in 3–6 months, touch enough stakeholders to build support, have a champion who’ll own them, and will teach the organisation something.</li><li><strong>Measure, refine, scale.</strong> Track results against the baseline from scoring. Use the lessons to move on to more Quick Wins, then Strategic Projects.</li></ol><h3>The bottom line</h3><p>AI agents are tools, not a strategy. They’re powerful in the right place and expensive in the wrong one.</p><p>The skill that separates successful agent programmes from pilot graveyards isn’t prompt engineering or picking the right model. It’s the discipline to say “not this one” and start with the boring, valuable task that’s ready to go.</p><p>Start small. Learn fast. Scale what works.</p><p><em>If you’ve shipped an agent that actually stuck, I’d love to hear what made it work. And if one flopped, which circle did it miss?</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=85b403384cf5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Most Companies Don’t Have a Tech Problem. They Have a Strategy Problem.]]></title>
            <link>https://medium.com/@leroyleowdev/most-companies-dont-have-a-tech-problem-they-have-a-strategy-problem-2a77d1c6f9b5?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/2a77d1c6f9b5</guid>
            <category><![CDATA[business-strategy]]></category>
            <category><![CDATA[digital-transformation]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 00:18:32 GMT</pubDate>
            <atom:updated>2026-10-09T00:38:03.329Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What Singapore’s IMDA–BCG Digital Acceleration Framework gets right about transformation, and what builders should take from it.</em></p><p><em>What Singapore’s IMDA–BCG Digital Acceleration Framework gets right about transformation, and what builders should take from it.</em></p><h3>The bottleneck moved, and most of us didn’t notice</h3><p>A few years ago, “we can’t afford the tech” was a fair excuse. It isn’t anymore.</p><p>Today a five-person company can rent AI assistants, process automation, BI dashboards and analytics as cloud services for less than a junior hire. Generative AI drafts the copy. Agentic AI runs the workflows. You can wire up a respectable digital stack in a few months.</p><p>So why do so many businesses still feel stuck?</p><p>I’ve been working through the Digital Acceleration Framework, built by Singapore’s Infocomm Media Development Authority (IMDA) and Boston Consulting Group (BCG). Its answer is blunt: <strong>technology and cost are no longer the binding constraint.</strong> What’s left is strategy clarity, the quality of your transformation plan, and whether your people are ready.</p><p>As a developer, that stung a little. We like to believe the next tool fixes things. This framework argues the tool is the easy part.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1ecuMJyVFoHiOhpBxIcdeA.png" /></figure><h3>Where most businesses actually stand</h3><p>In 2020, the IDC–Cisco Asia Pacific SMB Digital Maturity Study found that only <strong>16%</strong> of small and medium businesses had reached the advanced “Digital Challenger” or “Digital Native” stages. The other <strong>84%</strong> were “Digital Indifferent” (reacting to the market) or “Digital Observers” (chasing small efficiency wins).</p><p>Singapore, Japan and New Zealand led the region. But the pattern of low maturity wasn’t an Asia problem. It shows up in SMBs everywhere.</p><p>The study named two barriers, and neither was technology:</p><ul><li><strong>Skills.</strong> A shortage of digital talent, among leaders and staff alike, was the single biggest blocker.</li><li><strong>Culture.</strong> Leaders were wary of disrupting processes that worked. Employees feared for their jobs. Few companies rewarded experimentation.</li></ul><p>Then COVID hit, and <strong>69%</strong> of Asia-Pacific SMBs sped up their digitalisation. The firms further along the maturity curve saw up to 50% more sales, revenue and productivity than their peers. Digital Natives saw double the benefits.</p><p>Add Singapore’s fully funded digital reskilling programmes and the cost barrier mostly disappears. What remains is a planning and people problem.</p><h3>Six building blocks, one system</h3><p>The framework breaks transformation into six building blocks. The important idea isn’t the list. It’s that the blocks form <strong>one system</strong>, each feeding the next.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/973/1*lN7VSLn5S1ijTAXAAqFwDA.png" /></figure><p>Read it top down. Strategy decides where technology money goes. Technology only pays off when people can use it. Together they make new digital products and a digitised core possible. Ecosystem partnerships then amplify both.</p><p>Here’s each block with the principle the framework pins to it:</p><ul><li><strong>Business Strategy Driven by Digital.</strong> Name your goals (reach, efficiency, new revenue), then fund the digital work that serves them. <em>Digital transformation is a business strategy, not a technology project. Every digital initiative must trace back to a business objective.</em></li><li><strong>Leveraging the Power of Data &amp; Technology.</strong> Use AI, IoT and analytics to make better decisions, spot trends and optimise operations. <em>Tech is how you unlock the value already sitting in your data.</em></li><li><strong>Changing Ways of Working.</strong> Digital-ready leaders, upskilled staff, agile teams and governance that allows experiments. <em>People and culture multiply the tech spend; without them, tools sit unused.</em></li><li><strong>New Digital Growth.</strong> Build new digital products and services through internal teams, partnerships, acquisitions or start-up incubation, prototyping before you scale. <em>Digital transformation creates new value, not just operational efficiency. The highest returns come from new digital products, services, and business models.</em></li><li><strong>Digitalise the Core.</strong> Rework customer engagement, operations and support functions end to end. More on this below.</li><li><strong>Integrating Ecosystems.</strong> Share data and expertise with customers, suppliers and partners. <em>Mature digital businesses extend beyond their own walls, and network effects make that advantage compound.</em></li></ul><h3>Where the value actually lives: Digitalise the Core</h3><p>If you remember one block, make it this one. The framework says end-to-end digitalisation of the core is where most transformation value is created. It covers three areas.</p><p><strong>1. Customer offer and go-to-market.</strong> Redesign the whole customer journey: conversational marketing and sales, R&amp;D shaped by real-time customer data, micro-segment targeting, AI-assisted selling and dynamic pricing. The aim is one personalised experience across every touchpoint.</p><p><strong>2. Operations.</strong> Production, supply chain, procurement and service ops become connected systems. Think automation, IoT-driven predictive maintenance, full supply-chain visibility and AI-powered service. The aim is operations that flex with demand.</p><p><strong>3. Support functions.</strong> HR, finance, customer service and shared services go digital: AI-assisted recruitment, automated payroll, real-time financial analytics. The freed-up capacity moves to strategic work.</p><p>The line I keep coming back to: this means rethinking how each activity creates value, <em>not just automating what you already do</em>. Automating a broken approval chain gets you a faster broken approval chain.</p><h3>Measuring it: the Digital Acceleration Index</h3><p>A framework you can’t measure is a poster. IMDA and BCG pair this one with the <strong>Digital Acceleration Index (DAI)</strong>, a 0–100 score built from forty dimensions across strategy, technology, processes and people. The score drops you into one of four stages:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*o8OzEy1tQS9ARF4wJMLBKA.png" /></figure><p>Fully digitised products and operations; a platform ecosystem with network effects</p><p>The rule that matters: <strong>the stages are cumulative.</strong> You can’t leapfrog from Starter to Leader by buying an AI platform. Each stage stands on the one below it.</p><p>And the payoff is concrete. Per the IMDA–BCG research, Digital Leaders beat their less mature peers on every headline metric:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/724/1*p0VqMSrgXu_VDt5owPjsdA.png" /></figure><p>That’s a five-point gap on each line, roughly double the result. Not a rounding error.</p><h3>A builder’s honest take</h3><p>I like this framework. It puts strategy first, treats people as a multiplier rather than an afterthought, and gives you a score to argue about. But a few things are worth reading with a raised eyebrow.</p><p><strong>“Nearly double” is a bit loose.</strong> Cost reduction (9% vs 4%) is more than double. Revenue and enterprise value (11% vs 6%) are under. The direction is clear; the headline is rounded up.</p><p><strong>Correlation isn’t causation.</strong> Leaders outperform, but well-run, well-funded companies tend to be both more digital <em>and</em> more profitable. Digitising alone won’t guarantee the gap.</p><p><strong>“Tech is no longer a barrier” is an argument, not a measurement.</strong> The hard survey data is from 2020. The post-2024 claim about AI making maturity achievable in months is plausible, but it’s the authors’ read, not a study result. Anyone who has integrated a shiny AI tool into a 15-year-old ERP knows “seamless” is doing heavy lifting.</p><p><strong>The ladder is tidy; real companies aren’t.</strong> Plenty of firms are Performers in sales and Starters in finance. The forty-dimension DAI captures that better than a single stage label does, so look at the breakdown, not just the headline number.</p><p>None of this undermines the core message. It just means you should use the framework as a map, not a promise.</p><h3>Putting it to work</h3><p>If you’re a founder, a tech lead or the person everyone asks to “do the AI thing”, here’s how I’d use the framework on Monday:</p><ol><li><strong>Score yourself honestly.</strong> Run a maturity assessment and find your DAI stage. Benchmark against your sector, not against Big Tech.</li><li><strong>Write the business goal before the tech.</strong> Reach, efficiency or new revenue? Every initiative should trace back to one.</li><li><strong>Fix the floor of your current stage.</strong> Starters need cloud, ERP/CRM and a plan. Literates need skills, analytics and KPIs. Don’t skip ahead.</li><li><strong>Invest in people alongside tools.</strong> Budget upskilling and change management in the same line item as the software.</li><li><strong>Pick one core journey and redo it end to end.</strong> Rethink it, then automate it.</li><li><strong>Then look outward.</strong> New digital products and ecosystem partnerships pay off most once the core is solid.</li></ol><h3>The bottom line</h3><p>Digital transformation is a business strategy wearing a technology costume. The IMDA–BCG framework strips the costume off: six connected building blocks, a measurable ladder, and a clear payoff for climbing it.</p><p>The tools are cheap now. The thinking isn’t. That’s the work.</p><p><em>If this was useful, give it a clap and tell me in the comments: which stage is your company really at?</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2a77d1c6f9b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Find-Skill in Claude Code: How I Stopped Rebuilding Workflow from Stratch]]></title>
            <link>https://medium.com/@leroyleowdev/find-skill-in-claude-code-how-i-stopped-rebuilding-workflow-from-stratch-ab71047a84c1?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/ab71047a84c1</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[skills]]></category>
            <category><![CDATA[cursor-ai]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 12:53:08 GMT</pubDate>
            <atom:updated>2026-09-23T12:53:08.677Z</atom:updated>
            <content:encoded><![CDATA[<p>I hit a wall last month. I was deep in a multi-service TypeScript monorepo. Claude Code was already my daily driver — great at navigating the codebase, writing tests, and refactoring. But every time I needed something specialized (Docker multi-stage builds, a solid PR review checklist that caught the exact class of bugs, or a clean way to generate OpenAPI clients), I fell into the same loop:</p><p>1. Leave the terminal <br>2. Search GitHub / awesome-lists / random Medium posts <br>3. Clone a repo that *might* have a usable `SKILL.md` <br>4. Manually drop it into `~/.claude/skills/` <br>5. Restart Claude Code and hope the description was good enough for auto-activation</p><p>Half the time the skill was <strong>outdated</strong>, written for a <strong>different agent</strong>, or <strong>missing the frontmatter</strong> Claude Code expects. The other half I spent more time evaluating the skill than the original task would have taken.</p><p>That friction is what made me install <strong>find-skill</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R1ZzmtwzUL2ZsWkc" /><figcaption>Photo by <a href="https://unsplash.com/@berko?utm_source=medium&amp;utm_medium=referral">Ran Berkovich</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>What find-skill Actually Is</h3><p><a href="https://github.com/fockus/claude-skill-find-skill">find-skill </a>is a multi-agent skill discovery and installation tool. It gives you two slash commands inside Claude Code (and also works with Codex, OpenCode, and Cursor):</p><ul><li><em>/find-skill</em> — search a local catalogue of <strong>4,800+ skills</strong> from 14 sources</li><li><em>/install-skill</em> — install any of them with automatic format conversion for the target agent</li></ul><p>The catalogue is local-first (offline-capable), ranked by source trust + GitHub stars, and shared across all four agents. One update refreshes everything.</p><p>Sources include official Anthropic skills, the large skills.sh catalogue (Vercel-curated), several high-star awesome-lists, and a few community collections. Claude Code sees 100% of them because `SKILL.md` is its native format.</p><h3>Installation</h3><p>Pick whatever fits your workflow:</p><p><strong>Homebrew (macOS / Linux)</strong></p><pre><br>brew tap fockus/tap<br>brew install find-skill<br>find-skill # auto-detects Claude Code / Codex / OpenCode / Cursor<br></pre><p><strong>One-liner</strong></p><pre>curl -sSL https://raw.githubusercontent.com/fockus/claude-skill-find-skill/main/quick-install.sh | bas<br></pre><p><strong>Or target only Claude Code:</strong></p><pre>curl -sSL https://raw.githubusercontent.com/fockus/claude-skill-find-skill/main/quick-install.sh | bash -s - - target claud<br></pre><p><strong>Manual</strong></p><pre>git clone https://github.com/fockus/claude-skill-find-skill.git ~/.claude/skills/claude-skill-find-skill<br>cd ~/.claude/skills/claude-skill-find-skill<br>./install.sh</pre><p>After install you get:</p><ul><li>`~/.claude/skills/find-skill/SKILL.md` (the skill itself)</li><li>Shared cache at `~/.claude/skills/find-skill/cache/catalogue.json`</li><li>`/find-skill` and `/install-skill` available in Claude Code</li></ul><h3>Enter find-skill</h3><p>find-skill (from <a href="https://github.com/fockus/claude-skill-find-skill">fockus/claude-skill-find-skill</a>) is a meta-skill + CLI that:</p><ul><li>Maintains a local offline catalogue of <strong>4,835 skills</strong> from 14 sources</li><li>Ranks results by trust (official Anthropic first, then high-star awesome-lists, then community)</li><li>Filters by agent (Claude Code, Cursor, etc.)</li><li>Installs with automatic format conversion</li><li>Works fully offline after the first catalogue download</li></ul><p>It adds two slash commands to Claude Code:</p><ul><li>/find-skill</li><li>/install-skill</li></ul><h3>How to Use It</h3><p>Inside Claude Code, just type:</p><pre>/find-skill docker</pre><p>You’ll get the top 5 results ranked by relevance and source trust, filtered to Claude Code-compatible skills by default.</p><p>Useful flags:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/916/1*w9Wn9HoZYeROcUM8O3jmSw.png" /></figure><p>Once you see a skill you like:</p><pre>/install-skill owner/repo</pre><p>or simply by name if it’s in the catalogue:</p><pre>/install-skill memory-bank</pre><p>Advanced install options:</p><pre>/install-skill fockus/some-skill --target all          # Install for all 4 agents<br>/install-skill mono-repo --name specific-subskill      # Pick a sub-skill from a mono-repo<br>/install-skill https://github.com/user/skill-repo      # Full URL<br>/install-skill some-skill --force                      # Overwrite existing</pre><p>Inside any Claude Code session:</p><pre>/find-skill docker<br>/find-skill react - limit 10<br>/find-skill deploy - all<br>/find-skill python - page 2 - limit 5<br>/find-skill - top 20<br>/find-skill - stats</pre><p>The installer clones the repo, finds the SKILL.md (even deep inside mono-repos), converts the frontmatter if needed, and places everything in the right location.</p><h3>Keeping the Catalogue Fresh</h3><pre># Manual update<br>~/.claude/skills/find-skill/update-skills-catalogue.sh<br><br># Optional monthly cron<br>(crontab -l 2&gt;/dev/null; echo &quot;0 0 1 * * $HOME/.claude/skills/find-skill/update-skills-catalogue.sh&quot;) | crontab -</pre><h3>My Real Use Case and the Payoff</h3><p>I needed better Docker and deployment skills for a services-heavy repo. Previously I would have spent 20–40 minutes hunting, reading READMEs, and testing half-broken skills.</p><p>With find-skill I typed:</p><pre>/find-skill docker multi-stage<br></pre><p>Got a short ranked list (Anthropic and high-star community sources floated to the top). Installed the one that matched our patterns in one command. Ten minutes later Claude was generating Dockerfiles that already followed our internal conventions and suggested complementary skills for the deployment side.</p><p>Same story the next day with PR review skills and an OpenAPI client generator. The cognitive overhead of “is there already a skill for this?” basically disappeared. I stay in flow inside the terminal instead of context-switching to the browser.</p><p>Additional benefits I didn’t expect:</p><ul><li>Trust ranking surfaces the better-maintained skills first</li><li>Shared catalogue means the same search works if I hop over to Cursor or Codex</li><li>Local-first design means it still works on trains / restricted networks</li><li>Format conversion removes the “will this even load?” tax when trying skills written for other agents</li></ul><h3>When It’s Especially Useful</h3><ul><li>You keep pasting the same multi-step instructions into Claude</li><li>You’re exploring a new domain (security reviews, data pipelines, frontend design systems, etc.)</li><li>You work across multiple AI coding agents and want one discovery surface</li><li>You want to avoid reinventing workflows that the community has already packaged</li></ul><h3>Quick Tips</h3><ul><li>Keep the catalogue reasonably fresh — run the update script monthly or after big ecosystem releases.</li><li>Prefer skills from Anthropic, skills.sh, and the highest-star lists when quality matters.</li><li>After installing, glance at the `SKILL.md` description; a good description is what makes Claude auto-activate it later.</li><li>Use ` — dry-run` on `/install-skill` the first time you try something unfamiliar.</li></ul><p>If you’re already living inside Claude Code and occasionally think “someone must have already packaged this workflow,” install find-skill. It’s the closest thing we have right now to a package manager for agent skills — and it turned a recurring friction point into a two-second slash command for me.</p><p>Repo: <a href="https://github.com/fockus/claude-skill-find-skill">https://github.com/fockus/claude-skill-find-skill</a></p><p>Happy skill hunting.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ab71047a84c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Mastering Claude Code: Shortcuts, Sessions, and Checkpoints]]></title>
            <link>https://medium.com/@leroyleowdev/mastering-claude-code-shortcuts-sessions-and-checkpoints-2d092d5d450c?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/2d092d5d450c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[shortcuts]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Sun, 13 Sep 2026 01:52:19 GMT</pubDate>
            <atom:updated>2026-09-13T08:41:21.732Z</atom:updated>
            <content:encoded><![CDATA[<p>When building software with Claude Code, the real bottleneck is rarely generation speed; it is session drift and context poisoning. When an agent misunderstands a prompt, attempts an architectural refactor, and breaks your build, prompting your way out usually makes things worse. The agent remembers its own mistakes and layers bad fixes onto flawed foundations.</p><p>To work effectively, you must treat your session like an interactive scratchpad. Claude Code gives you built-in mechanisms to inspect state, checkpoint progress, and surgically revert file changes without touching your Git commit history.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AoV_zb4Vo2Vm8ebr" /><figcaption>Photo by <a href="https://unsplash.com/@dkfra19?utm_source=medium&amp;utm_medium=referral">Dimitri Karastelev</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Basic Claude Code Shortcuts</h3><p>Here are the essential Claude Code shortcuts for daily use. Keyboard behavior can vary slightly by operating system and terminal.</p><p>| `Esc` | Stop Claude’s current response or tool call</p><p>| `Esc` + `Esc` | Open the rewind menu when the input is empty</p><p>| `Ctrl+C` | Interrupt an operation or clear the current input</p><p>| `Ctrl+D` | Exit Claude Code</p><p>| `Ctrl+L` | Clear or redraw the terminal screen without deleting session history</p><p>| `Ctrl+R` | Search your command history</p><p>| `Ctrl+O` | Open the detailed transcript viewer</p><p>| `Ctrl+S` | Save the current prompt draft and restore it later</p><p>| `Ctrl+B` | Send a running Bash command or agent task to the background</p><p>| `Ctrl+G` | Open the current prompt in your default text editor</p><p>| `Shift+Tab` | Cycle through permission modes</p><p>| `Up` / `Down` | Move through previous prompts</p><p>| `Tab` | Accept an autocomplete suggestion</p><p>| `?` | Show the shortcut help panel when the input is empty</p><h3>Useful Input Shortcuts</h3><ul><li>`\` followed by `Enter` – Start a new line in a prompt.</li><li>`Option+Enter` on macOS – Insert a new line.</li><li>`Shift+Enter` – Insert a new line in supported terminals.</li><li>`Ctrl+J` – Insert a new line in any terminal.</li><li>`Ctrl+A` – Move to the beginning of the line.</li><li>`Ctrl+E` – Move to the end of the line.</li><li>`Ctrl+W` – Delete the previous word.</li><li>`Ctrl+U` – Delete from the cursor to the beginning of the line.</li><li>`Ctrl+K` – Delete from the cursor to the end of the line.</li></ul><h3>Basic Claude Code Prefixes</h3><p>| `/` | Run a slash command | `/help`</p><p>| `!` | Run a shell command | `! npm test`</p><p>| `@` | Reference a file or folder | `@src/app.ts`</p><h3>Essential Slash Command</h3><p>| `/help` | Show available commands</p><p>| `/clear` | Clear the current conversation</p><p>| `/compact` | Summarize the conversation to free context</p><p>| `/context` | Show context-window usage</p><p>| `/status` | Show session and configuration details</p><p>| `/model` | Change the active model</p><p>| `/resume` | Resume a previous session</p><p>| `/rewind` | Return to an earlier checkpoint</p><p>| `/diff` | Review current code changes</p><p>| `/init` | Create project instructions in `CLAUDE.md`</p><h3>Three Shortcuts to Learn First</h3><ol><li>`Esc` – Stop Claude when it takes the wrong direction.</li><li>`Esc` + `Esc`– Rewind code or conversation changes.</li><li>`Ctrl+R`– Quickly reuse an earlier command or prompt.</li></ol><p>You can also type `/` or press `?` inside Claude Code to see the commands and shortcuts available in your current installation. S</p><h3>The Command Center: Essential Session Shortcut</h3><p>The primary interface shortcuts you need to know:</p><ul><li>`Esc` `Esc`: The checkpoint shortcut. Pressing `Esc` twice rapidly when the prompt buffer is empty instantly launches the rewind and checkpoint browser. If you already have text typed in your input field, double-tapping `Esc` clears the prompt line and caches it into your history, letting you press `Up` later to recall it.</li><li>`<strong>/rewind</strong>` (or `/checkpoint`): Opens the interactive checkpoint menu if you prefer typing slash commands over hitting `Esc` `Esc`.</li><li>`<strong>/resume</strong>`: Allows you to inspect and switch back into past sessions across workspace directories.</li><li>`<strong>/clear</strong>`: Wipes the visible prompt screen. Note that on recent releases, `/clear` is no longer a permanent boundary – you can still invoke `/rewind` to reach points prior to the clear command.</li><li>`<strong>/compact</strong>`: Condenses running conversational context into an LLM-generated summary to clear memory headroom without altering any files on disk.</li></ul><h3>How Checkpoint Work Under The Hood</h3><p>Claude Code does not require manual save commands. By default, every single prompt that initiates a turn creates an internal snapshot of the working tree.</p><p>According to the official [<a href="https://code.claude.com/docs/en/checkpointing">Claude Code Checkpointing Documentation</a>], Claude Code retains file snapshots for the **100 most recent checkpoints** in a session, held across a retention sweep period of 30 days. These checkpoints persist across sessions, meaning if you resume an old session via `/resume`, the checkpoints remain intact.</p><h3>Reverting and Tracking Changes:The Checkpoint Menu</h3><p>When you trigger <strong>`Esc` `Esc`</strong> or type `<strong>/rewind</strong>`, Claude Code presents a chronological timeline of checkpoints labeled by timestamp, prompt instructions, and touched files. Once you select a checkpoint, you get granular control over how the rollback happens:</p><h3>Restore Checkpoint</h3><ol><li><strong>Restore code and conversation</strong>: Total rollback. It resets tracked files on disk back to their exact state at that moment and deletes all conversation that happened after it. Use this when the agent went completely off the rails.</li><li><strong>Restore conversation</strong>: Resets the conversation history back to that checkpoint, but leaves the current files on disk untouched. This is useful when the agent implemented the code correctly, but you want to purge conversational noise and explain the next phase with a clean context window.</li><li><strong>Restore code</strong>: Reverts file modifications back to the checkpoint while keeping your entire conversation history intact. This is ideal for debugging: the analysis, reasoning, and test logs stay visible in the chat, but the broken implementation gets wiped from your project tree.</li><li><strong>Summarize from here / Summarize up to here</strong>: Compresses the selected boundary into a summarized context block to save token window limits without modifying any files on disk.</li></ol><h3>Checkpoint Boundaries &amp; Gotchas</h3><p>Do not treat checkpoints as a replacement for Git. Checkpoints serve as transient, session-level safety nets. Be mindful of their hard technical boundaries:</p><ul><li><strong>Bash changes are invisible to checkpoints</strong>: If Claude executes `rm`, `mv`, or `cp` through terminal commands, or runs package installations like `npm install`, these file-system operations are <strong>not</strong> captured in file-tool checkpoints. Checkpoints only record direct edits made through Claude Code’s file editing APIs.</li><li><strong>Symlinks and hard links</strong>: Claude Code deliberately skips restoring symlinked or hard-linked files during a rewind to prevent corrupting external system files, displaying a warning instead.</li><li><strong>Background subagent edits</strong>: Background skills or disconnected worker runs are not rolled back via your primary session rewind menu; those require standard Git operations to revert.</li></ul><p>Keep sessions scoped to single features or discrete fixes, use `<strong>Esc</strong>` `<strong>Esc</strong>` aggressively whenever an agent takes a wrong turn, and commit working checkpoints directly to Git.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2d092d5d450c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Setting Up Remote Debugging in Visual Studio 2022: A Practical Guide]]></title>
            <link>https://medium.com/@leroyleowdev/setting-up-remote-debugging-in-visual-studio-2022-a-practical-guide-30b5a3fcb871?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/30b5a3fcb871</guid>
            <category><![CDATA[visual-studio]]></category>
            <category><![CDATA[debugging]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[dotnet]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 01:58:07 GMT</pubDate>
            <atom:updated>2026-08-03T01:58:07.398Z</atom:updated>
            <content:encoded><![CDATA[<p>Debugging issues in a production-like environment can be challenging when the application runs on a different machine than your development workstation. Fortunately, Visual Studio 2022 provides a robust remote debugging capability that allows developers to connect to and debug applications running on remote servers, virtual machines, or test environments without requiring Visual Studio to be installed on the target machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mCVe49oO-hKmT2Ae" /><figcaption>Photo by <a href="https://unsplash.com/@pankajpatel?utm_source=medium&amp;utm_medium=referral">Pankaj Patel</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In this guide, we’ll walk through the complete setup process for remote debugging in Visual Studio 2022.</p><h3>What Is Remote Debugging?</h3><p>Remote debugging enables Visual Studio to attach to and debug an application that is running on another computer. This is especially useful for:</p><ul><li>Troubleshooting applications deployed to test or staging servers.</li><li>Debugging applications running on virtual machines.</li><li>Investigating environment-specific issues.</li><li>Supporting distributed development and testing scenarios.</li></ul><h3>Prerequisites</h3><p>Before getting started, ensure the following requirements are met:</p><ul><li>Visual Studio 2022 is installed on your local development machine.</li><li>The target machine is accessible over the network.</li><li>You have sufficient permissions on the remote machine.</li><li>The Visual Studio Remote Tools are installed on the target machine.</li></ul><h3>Step 1: Install Remote Tools on the Target Machine</h3><p>The first step is to install the Visual Studio Remote Tools on the machine where the application will run.</p><h3>Installation Process</h3><ol><li>Download the latest <strong>Remote Tools for Visual Studio 2022</strong>.</li><li>Choose the installer that matches the architecture of the target machine: [x64, x86, ARM64]</li><li>Run the installer and complete the setup wizard.</li><li>Launch the <strong>Remote Debugger</strong> application (msvsmon.exe) after installation.</li></ol><p>Typical path for msvsmon.exe can be found in:</p><pre>C:\Program Files\Microsoft Visual Studio 17.0\Common7\IDE\Remote Debugger\x64 <br>or <br>C:\Program Files\Microsoft Visual Studio 17.0\Common7\IDE\Remote Debugger\x86</pre><blockquote><strong>Tip:</strong> Always use a Remote Tools version that is compatible with your Visual Studio installation. Version mismatches are among the most common causes of connection failures.</blockquote><h3>Step 2: Configure the Remote Debugger</h3><p>Once the Remote Debugger is running, it will display connection information similar to the following:</p><pre>Msvsmon started<br>Server Name: REMOTE-PC<br>User: DOMAIN\User</pre><p>During configuration:</p><ul><li>Verify the debugger is running under the correct user account.</li><li>Use <strong>Windows Authentication</strong> whenever possible.</li><li>Confirm the debugger is listening and ready for incoming connections.</li><li>Take note of the server name or IP address, as it will be required later.</li></ul><h3>Step 3: Configure Firewall Rules</h3><p>For Visual Studio to communicate with the remote debugger, network connectivity must be allowed between both machines.</p><p>Ensure that:</p><ul><li>msvsmon.exe is allowed through Windows Firewall.</li><li>Required debugger ports are accessible.</li><li>VPN or network security policies do not block communication.</li><li>Both machines can reach each other across the network.</li></ul><p>If connectivity issues occur, firewall restrictions should be one of the first areas to investigate.</p><h3>Step 4: Deploy and Run the Application</h3><p>Next, deploy your application to the remote machine.</p><p>Recommended steps:</p><ol><li>Build the application using the <strong>Debug</strong> configuration.</li><li>Copy or deploy the compiled application to the target machine.</li><li>Start the application before attempting to attach the debugger.</li></ol><p>Running a Debug build ensures that symbol information (.pdb files) is available, allowing breakpoints and variable inspection to function correctly.</p><h3>Step 5: Connect from Visual Studio 2022</h3><p>With the application running and the remote debugger active, switch to your local development machine.</p><h3>Attach to the Remote Process</h3><ol><li>Open your solution in Visual Studio 2022.</li><li>Navigate to:</li></ol><pre>Debug &gt; Attach to Process</pre><ol><li>In the <strong>Attach to Process</strong> dialog:</li></ol><ul><li>Set <strong>Connection Type</strong> to <strong>Remote</strong>.</li><li>Enter the remote server name or IP address in <strong>Connection Target</strong>.</li></ul><p>2. Click <strong>Refresh</strong> or <strong>Find</strong> to retrieve the list of available processes.</p><p>3. Select the appropriate process.</p><p>4. Click <strong>Attach</strong>.</p><p>At this point, Visual Studio establishes a debugging session with the application running on the remote machine.</p><h3>Step 6: Begin Debugging</h3><p>Once attached, you can use Visual Studio’s full debugging capabilities, including:</p><ul><li>Setting and hitting breakpoints.</li><li>Viewing local and global variables.</li><li>Monitoring watches.</li><li>Inspecting call stacks.</li><li>Using Step Into (F11) and Step Over (F10).</li><li>Evaluating expressions in real time.</li></ul><p>From a developer’s perspective, the experience is nearly identical to debugging a local application.</p><h3>Common Scenarios</h3><h3>Debugging an IIS-Hosted Web Application</h3><p>For ASP.NET applications running under IIS:</p><ol><li>Start the Remote Debugger on the web server.</li><li>Access the website once to ensure the IIS worker process starts.</li><li>In Visual Studio, attach to:</li></ol><pre>w3wp.exe</pre><p>This process hosts IIS application pools and is the most common debugging target for IIS deployments.</p><h3>Debugging a Windows Service</h3><p>For Windows Services, attach to the service executable:</p><pre>MyService.exe</pre><h3>Debugging a Console Application</h3><p>For standalone executables, attach directly to:</p><pre>MyApplication.exe</pre><h3>Troubleshooting Common Issues</h3><h3>Unable to Connect to the Remote Debugger</h3><p>Verify:</p><ul><li>Remote Debugger is running.</li><li>Firewall rules permit communication.</li><li>Authentication settings are correctly configured.</li><li>Network connectivity exists between both machines.</li></ul><h3>Process Not Visible in Visual Studio</h3><p>Check that:</p><ul><li>The application is actively running.</li><li>You are connected to the correct server.</li><li>The process architecture (x86/x64) matches expectations.</li><li>You have sufficient permissions to view and debug the process.</li></ul><h3>Breakpoints Are Not Being Hit</h3><p>Common causes include:</p><ul><li>Application deployed in Release mode.</li><li>Missing symbol (.pdb) files.</li><li>Source code does not match the deployed binaries.</li><li>Incorrect process attached.</li></ul><h3>Conclusion</h3><p>Remote debugging in Visual Studio 2022 is an invaluable feature for diagnosing issues that only occur outside the developer workstation. By installing the Visual Studio Remote Tools, configuring the Remote Debugger, ensuring network accessibility, and attaching to the correct process, developers can efficiently troubleshoot applications running on servers, virtual machines, or test environments.</p><p>When configured correctly, remote debugging provides the same powerful diagnostic capabilities as local debugging while allowing you to investigate issues in environments that closely resemble production.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30b5a3fcb871" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Supabase: Change In Policy For New Tables In Public Schema]]></title>
            <link>https://medium.com/@leroyleowdev/supabase-change-in-policy-for-new-tables-in-public-schema-44b431ca87f9?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/44b431ca87f9</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[supabase]]></category>
            <category><![CDATA[dotnet]]></category>
            <category><![CDATA[flutter]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Fri, 15 May 2026 23:55:02 GMT</pubDate>
            <atom:updated>2026-05-15T23:55:02.773Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>New tables in the public schema are no longer auto-exposed to the Data API Starting April 28, new Supabase projects can opt out of automatic Data API exposure for public schema tables. Explicit Postgres grants are now required to make a table reachable via PostgREST or GraphQL. This becomes the default for all new projects on May 30.</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U-m0mCSWzBa_Rpc0" /><figcaption>Photo by <a href="https://unsplash.com/@brett_jordan?utm_source=medium&amp;utm_medium=referral">Brett Jordan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>On 16th May 2026, I received an email from Supabase. If you’ve been building with Supabase, especially on the public schema, this is an important change you need to know about. Supabase has flipped the default behavior for newly created tables. They are no longer automatically exposed through the REST (PostgREST) and GraphQL APIs.</p><p>This maybe a big win from a security perspective but a subtle and meaningful shift in developer workflow</p><p>In the previous setup, you can creates a table like this:</p><pre>create table public.{table_name} (<br>  id uuid primary key default gen_random_uuid(),<br>  name text,<br>  location geography(Point, 4326)<br>);</pre><p>meant the table was <em>immediately</em> available at endpoints like:</p><p><a href="https://your-project.supabase.co/rest/v1/restaurants">https://your-project.supabase.co/rest/v1/</a>{table_name}</p><p>As long as your API keys (anon/authenticated roles) and Row Level Security (RLS) permitted it, anyone (or any service) with the URL could potentially interact with it.</p><p>This “convention over configuration” approach was convenient for rapid prototyping. But it was also dangerous.</p><pre>create table user_notes (<br>  id uuid primary key,<br>  user_id uuid,<br>  secret_notes text<br>);</pre><p>Forget to lock it down properly and boom — your sensitive data just became reachable via the public REST API. Many developers learned this the hard way.</p><h3>The New Reality: Explicit Exposure Required</h3><p>Now, when you create a table in the public schema, it exists in PostgreSQL but is invisible to Supabase’s auto-generated APIs until you explicitly grant access.</p><p>You have two main ways to expose it:</p><p><strong>via SQL</strong></p><pre>-- Grant read access (adjust as needed)<br>grant select, insert, update, delete <br>on table public.restaurants <br>to authenticated;<br><br>grant select <br>on table public.restaurants <br>to anon;  -- if you need public read</pre><p>Or through the Supabase Dashboard (UI-friendly for quick setups).</p><p>Without setting the above, you’ll likely encounter new errors after table creation:</p><ul><li>Could not find table in schema cache</li><li>permission denied for table restaurants</li></ul><p>Below is a post-creation template, if you have a tables created</p><pre>-- 1. Enable RLS (highly recommended)<br>alter table {table_name} enable row level security;<br><br>-- 2. Grant permissions to roles<br>grant select, insert, update, delete <br>on {table_name} to authenticated;<br><br>-- 3. Create sensible policies<br>create policy &quot;Anyone can view {table_name}&quot;<br>on {table_name}<br>for select<br>to authenticated<br>using (true);<br><br>create policy &quot;Users can add {table_name2}&quot; <br>on {table_name2}<br>for insert<br>to authenticated<br>with check (auth.uid() = user_id);</pre><h3>Why This Change Matters (And Why You Should Love It)</h3><p>This moves Supabase from a <strong>“everything is public unless you restrict it”</strong> model to <strong>“nothing is exposed unless you explicitly allow it”</strong>.</p><p>In an era of data breaches and compliance requirements (GDPR, anyone?), this is the right default. It forces intentionality around data exposure — especially valuable when your project grows to include tables like:</p><ul><li>payment_transactions</li><li>admin_logs</li><li>user_pii</li><li>Internal analytics tables</li></ul><h3>Final Thoughts</h3><p>This change reflects Supabase maturing from a delightful rapid-development tool into a more production-grade backend-as-a-service. The slight friction added during table creation is a small price to pay for better security posture.</p><p><strong>Pro tip:</strong> Add granting permissions and creating basic RLS policies to your “new table” checklist or automate it with migration scripts. Your future self (and your security auditor) will thank you.</p><p>The era of accidentally public tables is over. Good riddance.</p><p>What do you think — is this a welcome change or does it slow down your prototyping too much? Drop your thoughts in the comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44b431ca87f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ My Journey Preprocessing Data for Heart Attack Prediction]]></title>
            <link>https://medium.com/@leroyleowdev/my-journey-preprocessing-data-for-heart-attack-prediction-01353e180670?source=rss-85694527171------2</link>
            <guid isPermaLink="false">https://medium.com/p/01353e180670</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[process]]></category>
            <category><![CDATA[data-exploration]]></category>
            <dc:creator><![CDATA[Leroy Leow]]></dc:creator>
            <pubDate>Thu, 06 Nov 2025 13:56:51 GMT</pubDate>
            <atom:updated>2025-11-06T13:56:51.320Z</atom:updated>
            <content:encoded><![CDATA[<p>When I first started building a machine learning model to predict heart attacks, I quickly realized that raw data is rarely ready for training. It’s messy, inconsistent, and often biased. My goal was to create a model that could accurately predict heart attack risk based on patient data — things like cholesterol levels, resting blood pressure, age, and more. But before I could feed this data into any algorithm, I had to preprocess it carefully to ensure fairness, consistency, and performance.</p><p>I focused on techniques that help eliminate bias and prepare the data for ingestion, without diving into missing values or duplicate handling. Here are the four key preprocessing steps I used — and how they helped.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QI8RTnOilCHGowDi" /><figcaption>Photo by <a href="https://unsplash.com/@averey?utm_source=medium&amp;utm_medium=referral">Robina Weermeijer</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>🔘 Binarization</h3><p>Some features in my dataset were categorical but binary in nature. For example, the “exercise-induced angina” column had values like “yes” or “no.” To make this digestible for my model, I applied binarization — converting these values into 0s and 1s. So “yes” became 1 and “no” became 0. This helped the algorithm treat these features numerically without misinterpreting them as ordinal.</p><p><strong>Example:</strong> exercise_induced_angina = [&#39;yes&#39;, &#39;no&#39;, &#39;no&#39;, &#39;yes&#39;] → binarized = [1, 0, 0, 1]</p><h3>📉 Mean Removal</h3><p>Bias can creep in when features have vastly different means. For instance, age might have a mean around 50, while cholesterol could average 200. These differences can skew the model’s learning. I used mean removal to center each feature around zero by subtracting the mean from every value. This made the data more uniform and helped the model treat each feature with equal importance.</p><p><strong>Example:</strong> cholesterol = [180, 220, 200] Mean = 200 → mean_removed = [-20, 20, 0]</p><h3>📏 Scaling</h3><p>Next, I tackled the issue of feature magnitude. Resting blood pressure ranged from 90 to 180, while “number of major vessels” was just 0 to 3. Without scaling, the model might overemphasize the larger numbers. I used min-max scaling to bring all features into a common range, typically [0, 1].</p><p><strong>Example:</strong> resting_bp = [120, 150, 180] Min = 120, Max = 180 → scaled = [0.0, 0.5, 1.0]</p><h3>📊 Normalization</h3><p>Finally, I applied normalization to ensure that each data point had a consistent vector length. This was especially useful when using algorithms like k-NN or SVM, which rely on distance metrics. By normalizing each patient’s feature vector, I made sure that no single feature dominated due to its scale.</p><p><strong>Example:</strong> Patient vector = [50, 200, 120] → Normalize to unit length: normalized = [0.17, 0.69, 0.41] (approximate values)</p><h3>🧠 Final Thoughts</h3><p>Preprocessing isn’t just a technical step — it’s a critical part of building ethical, accurate models. By binarizing, removing means, scaling, and normalizing, I was able to reduce bias and give my heart attack prediction model a fair shot at learning from the data. If you’re diving into data science, especially in healthcare, mastering these techniques is a must.</p><pre>import pandas as pd<br>from sklearn.preprocessing import Binarizer, StandardScaler, MinMaxScaler, Normalizer<br><br># Sample dataset (simulated for illustration)<br>data = {<br>    &#39;age&#39;: [52, 58, 45, 63],<br>    &#39;cholesterol&#39;: [180, 220, 200, 240],<br>    &#39;resting_bp&#39;: [120, 150, 130, 160],<br>    &#39;exercise_induced_angina&#39;: [&#39;yes&#39;, &#39;no&#39;, &#39;no&#39;, &#39;yes&#39;]<br>}<br><br># Convert to DataFrame<br>df = pd.DataFrame(data)<br>print(&quot;Original Data:\n&quot;, df)<br><br># 1️⃣ Binarization: Convert &#39;yes&#39;/&#39;no&#39; to 1/0<br>df[&#39;exercise_induced_angina&#39;] = df[&#39;exercise_induced_angina&#39;].map({&#39;yes&#39;: 1, &#39;no&#39;: 0})<br>print(&quot;\nAfter Binarization:\n&quot;, df)<br><br># 2️⃣ Mean Removal: Center features around zero<br>scaler_mean = StandardScaler(with_std=False)  # Only remove mean, keep original scale<br>df_mean_removed = pd.DataFrame(scaler_mean.fit_transform(df), columns=df.columns)<br>print(&quot;\nAfter Mean Removal:\n&quot;, df_mean_removed)<br><br># 3️⃣ Scaling: Min-Max scaling to [0, 1]<br>scaler_minmax = MinMaxScaler()<br>df_scaled = pd.DataFrame(scaler_minmax.fit_transform(df), columns=df.columns)<br>print(&quot;\nAfter Min-Max Scaling:\n&quot;, df_scaled)<br><br># 4️⃣ Normalization: Normalize each row (patient vector)<br>normalizer = Normalizer()<br>df_normalized = pd.DataFrame(normalizer.fit_transform(df), columns=df.columns)<br>print(&quot;\nAfter Normalization:\n&quot;, df_normalized)</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=01353e180670" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>