用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/andrem-sec/psc-comet --skill skill-stocktake命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | skill-stocktake |
| description | Quality audit of installed skills with portfolio health dashboard |
| version | 0.1.0 |
| level | 3 |
| triggers | ["/skill-stocktake","audit skills","skill health check","review skill quality"] |
| context_files | ["context/project.md","context/learnings.md"] |
| steps | [{"name":"Mode Selection","description":"Choose Quick Scan (changed skills only) or Full Stocktake (all skills)"},{"name":"Skill Collection","description":"Gather all installed skills from .claude/skills/ directories"},{"name":"Evaluation","description":"Assess each skill against four dimensions using subagent"},{"name":"Verdict Assignment","description":"Assign one of four verdicts (Keep, Improve, Update, Retire, Merge)"},{"name":"Report Generation","description":"Generate portfolio health dashboard with findings"},{"name":"Cache Results","description":"Write results to ~/.claude/skills/skill-stocktake/results.json"}] |
Quality audit of your installed skills. Identifies skills that are outdated, redundant, or underperforming.
Without systematic skill audits, Claude:
The skill stocktake ensures your skill collection stays lean, current, and effective.
Duration: 5-10 minutes Scope: Only skills modified in the last 30 days or flagged in learnings.md Use When: Regular maintenance, after adding new skills, post-session cleanup
Process:
Duration: 20-30 minutes Scope: All skills in .claude/skills/core/ and .claude/skills/workflow/ Use When: Quarterly reviews, major version upgrades, before skill evolution
Process:
KEEP: Skill is current, actionable, unique, and effective. No changes needed.
IMPROVE: Skill has value but needs refinement. Issues might include:
UPDATE: Skill content is outdated. Triggers might include:
RETIRE: Skill provides no value or is harmful. Reasons to retire:
MERGE: Skill overlaps significantly with 1-2 other skills. Merge candidates when:
Actionability (High / Medium / Low)
Scope Fit (Right-Sized / Too Broad / Too Narrow)
Uniqueness (Unique / Overlaps / Redundant)
Currency (Current / Aging / Outdated)
When running Full Stocktake, spawn a subagent with this configuration:
subagent_type: researcher
tools: [Read, Grep, Glob]
prompt: |
Evaluate the skill at [skill-path].
Read the SKILL.md file and assess it against four dimensions:
1. Actionability: concrete, verifiable steps?
2. Scope Fit: single coherent problem, right-sized?
3. Uniqueness: overlaps with other skills?
4. Currency: uses current tools/practices?
Also read context/learnings.md and check if this skill appears in failure patterns.
Return a structured evaluation:
- Dimension scores (High/Medium/Low, etc.)
- Verdict (KEEP/IMPROVE/UPDATE/RETIRE/MERGE)
- Reasoning (2-3 sentences)
- If MERGE, list target skills
The subagent provides holistic AI judgment on subjective quality dimensions that resist simple rule-based checks.
Store stocktake results at ~/.claude/skills/skill-stocktake/results.json:
{
"timestamp": "2026-03-28T18:30:00Z",
"mode": "full",
"total_skills": 31,
"verdicts": {
"KEEP": 22,
"IMPROVE": 5,
"UPDATE": 2,
"RETIRE": 1,
"MERGE": 1
},
"skills": [
{
"name": "tdd",
"path": ".claude/skills/workflow/tdd/SKILL.md",
"verdict": "KEEP",
"actionability": "High",
"scope_fit": "Right-Sized",
"uniqueness": "Unique",
"currency": "Current",
"reasoning": "Well-defined Red-Green-Refactor phases, concrete coverage targets, no conflicts"
}
]
}
Quick Scan reads this cache to identify skills already evaluated recently.
Audit paralysis: Spending more time auditing skills than using them. Run Full Stocktake quarterly at most.
Over-retiring: Removing skills just because they're rarely used. Low-frequency high-value skills (like security-gate) should stay.
Subjective aesthetics: Retiring skills because "the writing style feels old." If the skill works and is current, keep it.
Ignoring merge recommendations: When two skills have 60%+ overlap, merging reduces context bloat and eliminates confusion about which to invoke.
Skipping the subagent: Evaluating all skills manually takes 2-3x longer and misses patterns. Use the subagent for holistic judgment.