Audit Project Skills and Suggest Updates

This skill audits a project's real recurring workflows before recommending skills, preferring evidence from memory, rollout summaries, existing skill folders, and current repo conventions over generic brainstorming. It maps the current project surface (repo guidance like AGENTS.md, README.md, roadmap/ledger docs), resolves the Codex memory base ($CODEX_HOME or ~/.codex), reads project history in order of specificity - starting with any in-context memory summary, then the MEMORY.md index searched by repo name/basename/cwd, then the 1-3 most relevant rollout summaries, falling back to raw session JSONL only when summaries lack a specific detail - and scans existing project-local skills (.agents/skills, .codex/skills, skills, including both SKILL.md and agents/openai.yaml) before suggesting anything new. It then compares project-local skills against recurring patterns (repeated validation sequences, failure shields, ownership boundaries, root-cause categories) to separate genuine "new skill" candidates from "update existing skill" candidates, and checks for overlap with global skills only after the project-local review.

Use it when the user asks what skills a project needs or which existing skills should be updated, or when recommendations should be grounded in project history, memory files, and local conventions rather than generic suggestions.

Its own failure shields apply as hard rules: do not invent recurring patterns without session or repo evidence, do not recommend duplicate skills when updating an existing one would suffice, do not rely on a single stale memory note if the repo has clearly evolved since, do not bulk-load all rollout summaries (stay targeted), do not skip rollout summaries and jump straight to raw sessions unless summaries are insufficient, do not recommend skills from themes alone (recommendations must come from repeated procedures, validation flows, or failure modes), and do not confuse a project's current implementation tasks with its reusable skill needs. If the user asks to actually create or update a recommended skill, this skill hands off to a $skill-creator skill rather than continuing the audit itself.

Input: a request to audit which skills a project needs, run against the current repo and its Codex memory/session history. Output: a compact audit with four sections - existing skills found and what each covers, suggested updates (skill name, why it's stale, highest-value change), suggested new skills (name, why, trigger, core workflow), and a priority-ranked recommendation order.

Output sections:
1. Existing skills
2. Suggested updates (skill name, why incomplete/stale, highest-value change)
3. Suggested new skills (name, why, trigger, core workflow)
4. Priority order

Reads from the Codex CLI memory system ($CODEX_HOME/memories/MEMORY.md, rollout_summaries/, sessions/), project-local skill folders (.agents/skills, .codex/skills, skills), and global skill locations ($CODEX_HOME/skills, $CODEX_HOME/skills/public). Hands off to a $skill-creator skill when the user wants a recommendation actually implemented.

Teams using Codex CLI-based agent workflows who want to periodically audit whether their project-local skills still match how the project is actually worked on, based on session evidence rather than guesswork.