We're Open Sourcing Our Agent Skills
The skills and hooks that make Claudette reliable are now open source.
Claude ignores skills about 30% of the time. We fixed that. Now we're releasing the fix.
The skills repo contains curated markdown prompts that define workflow rules. The hooks repo contains TypeScript automation that activates skills based on file patterns, commands, and context.
The Problem
Claude Code has a skill system. You write markdown files describing workflows. Claude is supposed to follow them.
It doesn't. Not reliably.
Skills get ignored when they conflict with Claude's default behavior. Skills get forgotten mid-task. Skills get partially applied. You tell Claude to write tests first, it writes code first. You tell Claude to fix the tool instead of working around it, it works around it anyway.
The skill system is powerful. But power without reliability is useless.
The Solution
Hooks activate skills automatically based on file patterns, commands, and context. When you edit a TypeScript file, the TypeScript review skill activates. When you run tests, the TDD skill activates. When you complete a feature, the dogfood skill activates.
This achieves 95% activation reliability. Claude doesn't forget the skill because the skill is injected at the right moment.
The Skills
tdd blocks implementation until tests exist. RED, then GREEN, then REFACTOR. No exceptions. The skill literally says "BLOCKED: Cannot proceed until test failure is proven." Claude takes this seriously.
dogfood forces you to use what you build. Building a CLI tool? Use it. Building an installer? Install with it. If the tool fails, you fix the tool. You don't work around it. Workarounds hide bugs that users will encounter.
differential-review performs security-focused code review. It calculates blast radius, checks test coverage, and models attackers. It adapts to codebase size: deep analysis for small repos, surgical analysis for large ones.
repo-hygiene handles housekeeping. Pre-work checks verify you're putting files in the right place. Documentation updates happen automatically. Writing style follows Paul Graham: short sentences, direct claims, no em dashes.
refactor-suggestions analyzes your changes and recommends improvements. It focuses on security, maintainability, readability, and functional programming purity. Suggestions are specific and actionable.
code-review-ts adds TypeScript-specific review guidelines. Strong types over runtime checks. No any without justification. Justified casts only. Memory leak detection for closures.
rick-rubin guards against scope creep. It's named after the producer because the approach is the same: remove everything that isn't essential. I wrote a separate post about this one.
Why Open Source
We built these skills for ourselves. They work. Now we want to see what others build with them.
The skills are opinionated. They enforce specific workflows. They reject common rationalizations. They block progress until requirements are met. This is the point. Agents that defer to users on technical decisions produce worse code.
If you build with AI, you need workflow enforcement. The prompts alone are not enough. You need structure. You need activation. You need skills that actually fire when they should.
The repos are live. Try them. Tell us what breaks.