agent-config

Audit AI agent configuration files (CLAUDE.md, AGENTS.md, preference files) for contradictions, duplication, bloat, and personal/team boundary violations. Use when unsure which instruction files are active for a project, after restructuring how your config is organised, or when the same rule turns up in two files with different wording.

Use before creating a CLAUDE.md, or before an edit that adds or rewrites a whole section of one — not for typo- or line-level fixes. Detects when a sibling AGENTS.md or README.md already covers the planned content, rewrites the CLAUDE.md as a pointer + Claude-only delta instead of a duplicate, and flags unactionable phrasing found in the sibling on the way through.

Use when the user asks "rate my prompt" / "evaluate my prompt(s)" / "how am I prompting?" / "give me prompt feedback" / "score my last prompt" / "4Ds analysis" / "prompt retrospective". Also use when the user asks for an end-of-session review of their work-with-Claude habits, OR after a session involving 20+ turns where the user explicitly wants improvement signal. NEVER fire unsolicited — prompt critique without consent is unwelcome.

Use before committing an addition to any prose-rule file — CLAUDE.md, AGENTS.md, a docs/ page, a README, or a PR template. Fires when a diff ADDS conventions, rules or procedures in prose. Auto-trigger after writing such an addition and before `git commit`, and when the user says a file is "too big", "has duplicates", or asks whether a change "is really necessary". Skip for code, tests, config values, pure deletions, rewording that adds no rule, and new files with no sibling docs. For the CLAUDE.md-vs-sibling file-structure rewrite specifically, defer to `claude-md-pointer-check`; this skill covers rule-level dedup across any prose-rule file.

Use ONLY when the user explicitly wants to compact a long session without losing its learnings — triggers like "save before compact" / "wrap up and compact" / "session's too long, compact but keep the takeaways" / "capture learnings then compact". NEVER fire unsolicited. This skill APPLIES changes (CLAUDE.md / preference-file additions, memory, new skills) and writes a resume brief, then stops for the user to run /compact — it never runs /compact itself. Thin sessions get only the resume brief; in-flight high-stakes ops defer (Step 1 has the thresholds).

Use ONLY when the user explicitly asks for an end-of-session retrospective — triggers like "session retro" / "end of session retro" / "session retrospective" / "wrap up the session" / "what did we learn this session" / "what should I take away from this session" / "let's reflect on this session". NEVER fire unsolicited — this skill produces evaluative feedback on the user's prompting, and reflection without consent is unwelcome. Skip when the session has fewer than ~15 substantive turns (signal too thin), or when the user is mid-incident / mid-deploy. If a sub-skill already ran this session, the retro asks whether to re-run or reuse its output rather than silently duplicating it.

Use when you notice the user repeating the same instruction / asking the same kind of question / correcting you on the same kind of mistake / doing the same manual operation three or more times in a session. Also trigger when the user explicitly asks "what should be a skill?" / "what am I doing manually?" / "where am I repeating myself?". Skip during the first ~5 turns of a session — patterns need history to be observable.

Use when a portable or generic skill has never been executed against a real project, before trusting it or calling it done. Fire on "validate the skill", "does this skill actually work", "test it against a real repo", when a skill claims to work on "any <stack> project" and no run is on record, or when a skill was reviewed by reading it rather than running it. Skip for project-specific skills, which are allowed to assume one build's wiring.