メニュー
Use when auditing the health of Cartographer atlases across all tracked repos. Reports fresh-atlas coverage, staleness distribution, atlas size, and layer population. Triggers on "audit atlas", "atlas health", "cartographer coverage".
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use when designing a goal-oriented agent loop or reviewing one for runaway risk - build gate, machine-decidable goal, anti-Goodhart boundaries, five failure modes. Triggers on "design a loop", "review this loop", "will this loop run away".
Use when running OpenAI Codex via the codex CLI (copilot CLI fallback) for code analysis, refactoring, or editing. Triggers on "run codex", "codex analyze", "ask codex", "copilot with codex".
Use when running Google Gemini via the native Gemini CLI for code analysis, refactoring, or editing. Triggers on "run gemini", "gemini analyze", "ask gemini".
Use when running a local Qwen model (via llama.cpp and the pi agent) for code analysis, refactoring, or editing - local inference, no API cost. Triggers on "run qwen", "qwen analyze", "ask qwen", "local model task".
Use when running Anthropic Claude Sonnet via the native claude CLI (headless) for code analysis, refactoring, or editing. Triggers on "run sonnet", "sonnet analyze", "ask sonnet".
Use when converting a plan, design doc, brainstorming output, or discovery doc into a structured PRD saved to the backlog. Triggers on "create PRD", "create PRD from plan", "make PRD", "convert to PRD", "save plan as PRD".
| name | audit-atlas |
| description | Use when auditing the health of Cartographer atlases across all tracked repos. Reports fresh-atlas coverage, staleness distribution, atlas size, and layer population. Triggers on "audit atlas", "atlas health", "cartographer coverage". |
Report on atlas coverage, freshness, and content quality across
~/.claude/cartographer/projects/. The aggregation is deterministic, so a
script computes it (PRD 00046) — never scan the atlas files in-model.
python3 ~/.claude/skills/audit-atlas/scripts/report.py
It scans every projects/<hash>/atlas.json (cross-referencing
~/.claude/instincts/projects.json for repos active in the last 30 days) and
prints four sections: Fresh-atlas coverage, Staleness distribution,
Atlas size, Layer population, plus a malformed: count. Fresh = no
staleness.flag AND surveyed under 14 days ago (the flag is authoritative —
the cartographer-stop hook computes the full 50-commit/14-day check with
per-repo overrides). An empty projects dir prints zero counts — that is the
LOW: no atlases found finding.
active_fresh_pct below 80% (the PRD target: >80% of
actively-edited repos have a fresh atlas), and any atlas.md listed over
the 5KB budget.under-enriched atlases (empty naming/error_style/
forbidden_imports/dependency_edges — re-run /survey to enrich) and each
degraded atlas (tree-sitter fallback; re-survey once tree-sitter is
installed), plus the no-atlases case.Synthesize a ## Findings section (CRITICAL/HIGH/MEDIUM/LOW) followed by a
## Summary block quoting the script's counts verbatim: total atlases,
fresh/stale, active_fresh_pct, degraded count, and the median sizes.
~/.claude/cartographer/projects/<hash>/; schema keys:
head_sha (absent for non-git dirs), surveyed_at (ISO-8601 UTC),
layers, forbidden_imports, naming, error_style,
dependency_edges; optional staleness, truncated, degraded.cd <repo_path> then /survey --refresh.