Audit all the instructions together
context-audit tells an agent to inventory the guidance it receives from
AGENTS.md, CLAUDE.md, skills, hooks, tool descriptions, and harness prompts.
It reads those layers together, then classifies rules as conflicts, duplicates,
obvious guidance, outdated blanket constraints, or project-specific gotchas.
Use it when an agent ignores rules or a project instruction file keeps growing. The output is a proposed diff with before-and-after line counts, not an automatic rewrite.
Keep the human approval step
The skill explicitly says to propose deletions rather than apply them. That matters because an awkward rule may be scar tissue from a real production incident. Ask the agent to quote both sides of every conflict and flag the one deletion it is least confident about, then approve or reject each change.
This is a community skill distilled from public context-engineering ideas, not an official Anthropic tool. Installation behaviour also differs by agent; the repository documents those differences.