René Dekkers

The review gate opens when the tests are green. Those tests were written by the same AI, in the same session, with the implementation in context — so they agree with the code instead of the requirement. Here’s how to tell, and what actually fixes it.

A shared skill and a team plugin make everyone fast. That’s the problem this post is about: when AI writes the code and a human clicks approve, who actually understood it? Two review gates — an automated one and a human one — keep the developer, not the AI, accountable.

In part 1 you turned a convention into a skill and committed it. This post is about the next problem: your real setup is a skill plus a review subagent plus a build hook plus an MCP server — and every developer needs the same one. A plugin packages all of it and ships it with one install command.

Every team has conventions that live in one person’s head and spread by code review. When you use Claude Code, that knowledge quietly forks: your setup knows your rules, nobody else’s does. Skills put the convention in the repo, so the whole team — and their AI — works the same way.

A lot of code is hard to test not because the business logic is complex, but because decisions and side effects are woven together. Functional Core, Imperative Shell is a simple rule for separating the two — and I pointed Claude Code at a tangled handler to do the split.

An order model with a status enum and a handful of nullable fields can represent far more states than are legal. I pointed Claude Code at one, had it count the illegal combinations, and let it redesign the type so they can’t be constructed at all.

One method returned a bool; another trusted that you’d called it first. That gap was the root of five separate problems. Here’s the refactor — parse, don’t validate — and how Claude Code did the mechanical work.

You get a feature request. Maybe a ticket, maybe just a Slack message. You open your IDE, create a branch, and start building. Claude Code writes the code, you review it, things move fast. Then, three hours in, you discover you’ve made five architectural decisions you never consciously made. The subscription entity is attached to the User because Claude assumed that. The tier check happens in the controller because that’s where it fit. There’s no audit trail because you never said you needed one. ...

Boris Cherny ships 10–30 pull requests a day and writes essentially zero lines of code by hand. That number spread across developer Twitter in January 2026, and since then I’ve watched teams respond in the worst possible way: spin up ten Claude sessions, produce ten times the chaos. The workflow matters. What makes it work is not the session count. Why this is worth your attention Cherny created Claude Code and leads it at Anthropic. Before that he was a Principal Engineer at Meta. He is not a productivity influencer who dabbled with AI for a week — he built the tool, uses it at extreme volume, and has been candid about what actually drives the numbers. ...

Eight days after publishing a guide to tuning Claude Code, a single week of releases — Claude Fable 5, a one-flag safe mode, Stop hooks that talk back, and guardrails that used to be your job — forced a re-tune. Here’s the delta, mapped onto the same layers, and what we changed in our .NET codebase.