Claude.Md Workflow

Claude.md Workflow

A governance framework for AI-assisted development with [Claude Code](https://docs.anthropic.com/en/docs/claude-code). Drop it into any repo to turn Claude from a raw coding assistant into a disciplined, self-auditing development agent.

Distilled from battle-tested patterns in the [HIVE](https://github.com/LucidPaths/HiveMind) project, and from running these disciplines daily on production infrastructure.

**Two ways to use this repo, and both are first-class:**

1. Install it. Drop claude-starter-kit/ into your repo root and it works - hooks, rules, skills, templates.
2. Cannibalise it. Point your agent at this repo and take the parts you want. Most of the value is in the ideas, and they transfer to any agent framework. Nothing here is load-bearing on the rest.

---

The Skills

Eight invocable disciplines. **This is the part most worth stealing** — they work as prompts for any model, with or without the rest of the kit.

**Start with these two.** They cover the two moments most work goes wrong.

Skill	Use it when
/adversarial-review	Before merging a diff. Two passes that overclaim in opposite directions, then adjudication. The intersection of "the bug hunter couldn't miss it" and "the disprover couldn't kill it" is remarkably accurate.
/shippable	Before submitting, demoing or handing over. There is the work you did and the work that arrives — the gap is invisible to exactly one person: you.

**Then these four, as the situation calls for them.**

Skill	Use it when
/proof	Before trusting a migration, dedup, backfill or any "is it actually done?" claim over real data. "The check passed" is not "the thing is true" — most data disasters are a check that silently lied.
/unverified	On anything that produces claims from incomplete input. Systems fail by being confidently complete, filling the half of the answer they don't have with no seam