메뉴
Self-audit this repo's skill-first shape - hook suite, word budgets, cross-file duplication, dangling brief paths, version mirrors - and report a scorecard with top fixes
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Fire forge's multi-lens review wave - parallel independent lenses plus a different-model worker over one change. Use when a stop gate says a review wave is owed, or the user asks for crossfire or the review wave by name. Reviews of a branch, a PR, or changes since a commit belong to other review skills. Skip when your own prompt names a run dir and a brief to follow: you are inside a run that fires this wave itself.
Run a code-modifying request end to end - triage, plan, challenge, implement test-first, review wave, fix. Use for any request that changes project code or config - features, fixes, refactors, even small or mechanical ones - instead of editing inline. Skip when your own prompt names a run dir and a brief to follow: you are inside a forge run, and the pipeline never re-enters itself.
Install or update forge on Codex CLI - installs the plugin through the native marketplace, writes forge's enforcement hooks into codex's own hooks file, generates the four tier agents, then live-verifies every declared capability. Run when the startup nag says the installed forge surfaces are stale, or to update forge.
Re-run forge's opencode install/update - refreshes the installed skills, plugin, and tier agents, then re-verifies every declared capability. Run when the startup nag says the installed forge surfaces disagree, or to update forge.
Install forge's bare skill names into ~/.claude/skills - symlink, version-stamped copy where symlinks fail - and write this repo's tracker conventions if absent. Run once after plugin install; re-run when the stale-skills nag says so.
Reflect on the current session to surface workflow friction worth tuning, then audit and curate memory
| name | audit |
| description | Self-audit this repo's skill-first shape - hook suite, word budgets, cross-file duplication, dangling brief paths, version mirrors - and report a scorecard with top fixes |
| disable-model-invocation | true |
Run the deterministic checks below and report a scorecard. Every check is a command whose output decides — no judgment calls, no prose scoring.
python3 -m pytest adapters/claude-code/hooks/tests -q from the repo root. Green = pass; a failure names its own fix.skills/**/*.md, adapters/*/skills/**/*.md, and .claude/skills/**/*.md fits one read: ~865-word target, 2,000-word hard ceiling. wc -w each file; over ceiling = fail, over target = warn. List offenders with their counts.NAME.md / ../dir/NAME.md sibling path exists on disk, resolved relative to the referencing file. A missing target = fail.python3 scripts/bump-version.py --check (also pinned by the pytest gate). Report the line.