Claude MD Generator

Most CLAUDE.md files fail one of two ways. They're skipped entirely — and the user wonders why Claude Code behaves like a generic assistant. Or they're a 300-line wall of generic boilerplate that Claude partially ignores, because attention degrades on long files and aspirational instructions ("write clean code") aren't testable. This skill produces files that avoid both failure modes — short, opinionated, and built around the 20% of instructions that drive 80% of behaviour change.

It runs a short interview rather than handing you a template. Round one covers role, project, and team shape. Round two captures working style — and crucially the "what annoys you about AI assistants?" question, because pet peeves map directly to the most valuable negative instructions in the file ("Do NOT add inline comments unless the logic is genuinely non-obvious"). Rounds three and four cover technical or domain specifics and the tasks you'll use Claude for most. The skill skips irrelevant rounds and, if you provide everything upfront, skips the interview entirely.

The generation rules are strict. Every instruction must be testable — you could verify whether Claude followed it. No "be helpful" or "write clean code". Total length stays under 150 lines, because Claude's attention to CLAUDE.md fades past that; overflow goes into a separate docs/claude-extended.md that Claude can read on demand. Sections match the role — a PM doesn't get a Code Style section, a developer doesn't get a Writing Style section. And the Common Tasks block — the highest-ROI section for non-developers — gets explicit When I say "X" → do Y templates rather than vague capability lists.

When users don't know what they want yet, the skill doesn't stall. It asks two questions (role, primary use case), generates a sensible starter file based on role defaults, and marks the customisable sections with <!-- CUSTOMIZE: ... --> comments — then tells you to revisit it after a week of use, when you'll actually know what annoys you. It also recognises when the real task is an audit, not a generation, and switches modes rather than pretending to start from scratch.