Five reads this time on the practical edges of working with AI coding tools, where the hype meets the reality of shipping.
Mattias Geniar’s How We Implemented a Mobile-Friendly Oh Dear UI with AI is the most honest case study I’ve read on the AI-assisted development ratio. They used Chief (a tool that breaks projects into tasks and runs Claude Code in a loop) to generate 21 commits touching 226 files in a single evening, adding responsive Tailwind classes, wrapping tables in scroll containers, stacking layouts vertically. Structurally correct. But it didn’t feel right. When your site goes down at 3am and you’re checking alerts on your phone, you need specific information immediately, and the AI didn’t know what mattered most in each context. That’s domain knowledge, built from years of using the product. The AI got them 60% of the way; the remaining 40% took three weeks and three developers. His summary is spot-on: AI is great at the tedious parts, bad at the creative parts. The scaffold is the easy part. The house is the hard part.
Birgitta Böckeler’s Humans and Agents on Martin Fowler’s blog explores the emerging collaboration patterns between human developers and coding agents. She identifies the key question: how do you keep the human meaningfully in the loop when the agent is doing most of the execution? The answer isn’t just code review, it’s about staying close enough to the agent’s reasoning that you can steer before things go sideways, not just inspect after. The piece pairs naturally with the Oh Dear post as a higher-level framing of the same problem.
Why We Wrote Our Own SEO Plugin from the My Buddy Claude project is the kind of pragmatic build log I enjoy. Instead of installing Yoast (10,000–50,000 lines of code, 8–15 admin pages, 20–100ms per page load), they built a 463-line mu-plugin that covers exactly four things: meta descriptions, Open Graph tags, Twitter cards, and JSON-LD structured data. It runs in 2ms. The architecture is clean, one function gathers all SEO context, four consumers render their specific tags. WordPress handles sitemaps natively since 5.5. If your site doesn’t need readability scores and keyword density analysis, this is the approach worth stealing.
Simon Willison’s Agentic Manual Testing proposes using coding agents as exploratory testers, instruct them to navigate your app like a real user, try edge cases, find broken flows. It doesn’t replace your test suite; it sits on top of it. The key insight is that agents are tireless and thorough in ways humans aren’t, they’ll try the 47th variation of an input without getting bored. Worth experimenting with on any project that has a web UI.
HumanLayer’s Writing a Good CLAUDE.md is the most research-backed guide I’ve found on the topic. The core findings: Claude Code’s system prompt already consumes ~50 of the ~150–200 instructions a frontier model can reliably follow. Every instruction you add to CLAUDE.md competes with those, and as instruction count increases, following quality decreases uniformly, it doesn’t just ignore the newest ones, it starts ignoring all of them. Their recommendation: keep CLAUDE.md under 60 lines (theirs is), make everything universally applicable, and use progressive disclosure, point Claude to separate docs files it can read when relevant rather than stuffing everything into one place. The “less is more” finding is counterintuitive and well-supported.