<?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 Станка Далекова on Medium]]></title>
        <description><![CDATA[Stories by Станка Далекова on Medium]]></description>
        <link>https://medium.com/@stanska?source=rss-82c6ebb7cb6e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*g9tEG_MyBs71lBwS.</url>
            <title>Stories by Станка Далекова on Medium</title>
            <link>https://medium.com/@stanska?source=rss-82c6ebb7cb6e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 17:20:46 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@stanska/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[Uncle Bob Agent Skills]]></title>
            <link>https://medium.com/@stanska/we-turned-uncle-bob-into-a-plugin-e02e150c7a22?source=rss-82c6ebb7cb6e------2</link>
            <guid isPermaLink="false">https://medium.com/p/e02e150c7a22</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-coding]]></category>
            <category><![CDATA[agent-skills]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Станка Далекова]]></dc:creator>
            <pubDate>Tue, 14 Jul 2026 19:59:21 GMT</pubDate>
            <atom:updated>2026-07-15T08:36:00.722Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/911/1*kNbtWxGpGy8EQQRfkpP_5A.jpeg" /></figure><p><em>How two twenty-year-old books became an enforcement mechanism for AI coding agents.</em></p><p>AI coding agents write code fast, but technical debt is accumulating just as fast. We tried rules in CLAUDE.md: keep functions small, don’t mix layers, name things after the domain. But CLAUDE.md is always loaded into the context, so we have to be careful what goes in it.</p><p>What we wanted was a software engineering practices checklist the coding agent has to follow.</p><h3>Agent Skills</h3><p>Agent Skills are a standard introduced by Anthropic in October 2025 and quickly adopted by other agents like Codex, Cursor, and Goose. A skill is a markdown file with a description and a body of instructions. The description, part of the meta information, is always loaded in the context, but the actual instructions — along with any helper scripts, reference files, and domain-specific knowledge a skill may carry — are progressively disclosed. Skills can be triggered explicitly by invoking them (/uncle-bob:clean-code) or automatically when a task matches — &quot;refactor this&quot;, &quot;clean this up&quot;.</p><p>Skill files are versioned, reviewable in a PR, and shareable through a plugin marketplace. Improve the checklist once, and everyone’s agent gets better on the next install.</p><p>The books I always recommend for becoming a good software engineer are Bob Martin’s <em>Clean Code</em> and <em>Clean Architecture</em>. That’s where I found the most battle-tested code-quality checklist that already exists.</p><p>We built a plugin called uncle-bob. Two skills, one per book.</p><p>Creating a skill is easy using the meta skill /skill-creator.</p><h3>What’s actually in it</h3><p>The books make arguments; the skill keeps only the conclusions. Every line had to pass one test: can an agent check this against a diff?</p><p>So the clean-code skill isn’t an essay about naming. It’s this:</p><pre>Names<br>- Names describe **what**, not **how**: `fetchUserById` not `doUserQuery`<br>- Boolean names start with `is`, `has`, `can`, `should`<br>- No type encoding in names (`userList` → `users`)</pre><pre>Functions<br>- **One job per function.** If you need &quot;and&quot; to describe it, split it.<br>- Max ~20 lines is a smell; max 3 arguments is the target<br>- No flag arguments - split into two functions instead<br>- Return early to avoid deep nesting; avoid `else` after a `return`</pre><pre>Error handling<br>- Use exceptions, not return codes or sentinel values<br>- Never return null from a function that should return a collection - return empty</pre><p>The clean-architecture skill does the same for structure, built around the Dependency Rule stated as an absolute:</p><pre>Source code dependencies point inward only.</pre><pre>Frameworks &amp; Drivers → Interface Adapters → Use Cases → Entities</pre><p>One instruction mattered more than any rule: apply the checklist <em>while making changes</em>, not as a cleanup pass. An agent that writes messy code and tidies up later burns tokens and sometimes decides the cleanup isn’t worth it. Same reason experienced engineers don’t write sloppy drafts to fix later — later has a way of not arriving.</p><h3>Running it on existing code</h3><p>You can’t ship a code-quality tool without pointing it at yourself. I ran the architecture skill on one existing module — an MCP server, a few thousand lines of Python grown the way proof-of-concepts grow: fast, under deadline.</p><p>Two minutes later it produced a report. The report was rude.</p><p><strong>Finding one:</strong> the server file was a God Object. Nine hundred lines mixing framework wiring with pure domain logic that had <em>zero</em> framework dependencies — it had just accumulated there, one PR at a time, because that’s where the cursor happened to be. Suggested fix, appropriately minimal: two files, not seven layers.</p><p><strong>Finding two:</strong> a client module named backend_client.py — after our infrastructure, not what it talks to. The name describes our architecture diagram, not the domain, and it goes stale the day the backend stops being one. Renamed to what it is: a payment API client.</p><p><strong>Finding three:</strong> two files with the same name in sibling packages, one importing from the other. A maintenance trap we’d all stopped seeing.</p><p><strong>Finding four,</strong> my favorite for its pettiness: a six-line wrapper re-exporting one function under the same signature. Pure ceremony. Verdict: delete it. We deleted it.</p><p>None of these are exotic. That’s the point. They’re the gradual erosion human reviewers stop noticing because each PR looked fine. The agent doesn’t know things we don’t — it applies the <em>whole</em> checklist, <em>every</em> time, with no fatigue and no diplomatic instinct to let the small stuff slide.</p><h3>Fast doesn’t mean dirty</h3><p>Writing fast is no excuse for abandoning clean code. The agent has seen every pattern in existence — the good, the bad, and the 900-line God Object — so without explicit direction, you get a coin flip.</p><h3>Where this goes</h3><p>Any standard your team holds can become a skill. Your REST conventions. Your security review checklist. Your rules about migrations, logging, date libraries. Right now they live in wiki pages nobody reads and the heads of two senior engineers who will eventually change jobs. Each could be a markdown file that loads at the exact moment the agent is about to do the relevant thing.</p><p>Writing one takes an afternoon — and don’t write skills by hand, use the skill-creation meta skill.</p><p>There’s a mild irony here. Uncle Bob wrote those books to change how people <em>think</em>, argument by argument. We stripped the arguments and kept the conclusions, for a reader that doesn’t need convincing — just telling.</p><p>But principles from 2008, written for humans reviewing humans, are now running inside the loop of a machine writing code at machine speed — and catching the exact same sins. The medium changed completely. The sins didn’t.</p><p>Standards used to live in books, then wikis, then review comments. Now they can live in the loop.</p><p><em>The plugin is two markdown files and a manifest — nothing about it is specific to us. It’s available at </em><a href="https://github.com/stanska/claude-code-marketplace/tree/main/plugins/uncle-bob"><em>github.com/stanska/claude-code-marketplace</em></a><em>, installable via the Claude Code plugin marketplace with </em><em>/plugin install uncle-bob. Skills adapted from Clean Code (2008) and Clean Architecture (2017) by Robert C. Martin.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e02e150c7a22" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>