claude-craft

Two halves.

The design skills answer questions rather than restating API signatures: how should the loop terminate, when does delegating to a subagent actually pay for itself, what belongs in a tool's schema, where should an invariant be enforced, how do you make an agent's output parse reliably when it feeds code rather than a person, and what to do about a rule the model keeps violating despite being told.

The audit skills point the same rules at a repository. config-audit is the one-shot health check — it runs the skills, agents and hooks audits in parallel, adds a settings and permissions pass, and merges everything into one ranked report. Below it sit the individual passes: review-skills for a tree of SKILL.md files, review-agents for a roster of subagents (including whether an agent should have been a skill), review-hooks for hook scripts together with their settings registrations, and agent-sdk-review for a codebase that uses the SDK directly.

It also carries the workflow skills that sit around that work: debug for a debugging session run end to end, repo-walkthrough for orienting in an unfamiliar codebase, and handoff for writing a self-contained handoff.

Reach for it when you are building agentic software rather than using it, or when a Claude Code configuration has grown past the point where anyone can say what is in it.