Skills-Summarize-Audit — Skill Library & Project Profile-Driven Audit (Linked with summarize)
Perform multi-dimensional profiling, 7-dimensional scoring, capacity analysis, and optimization recommendations for your skill library or current project, ultimately outputting an actionable audit report.
It helps you:
- Understand the health of your skill library (🟢/🟡/🔴) and trend changes (📈 ↑/↓)
- Discover zombie skills and redundant overlaps
- Calculate how many more skills an Agent can install (multi-factor effective capacity warnings)
- Get an actionable list of archive/install/update operations
- Roll back after mistakes with a single command (
--undo)- Integrate into CI/CD pipelines (JSON output + gating logic)
Trigger Words
This skill uses standalone word triggering: trigger words must be sent independently (optionally with subcommands) and will NOT trigger in the middle of a sentence. Designed to work with the summarize skill; 技能总结 can directly trigger an audit.
- Skill Audit:
skills-audit,skill-audit,skill-check,capability-audit,tool-audit,skills-summarize-audit - Project Audit:
project-audit,project-check,project-diagnosis,project-profile,env-audit - Undo Operations:
undo-snapshot,rollback,revert - Subcommands:
deep,recommend,profile,health,quick,ci/--ci,undo/--undo
Trigger Matrix
| Mode | English Trigger | Immediate Action |
|---|---|---|
| Skill Audit | skills-audit, skill-audit, skill-check, capability-audit, tool-audit, skills-summarize-audit | Full audit of skill library |
| Project Audit | project-audit, project-check, project-diagnosis, project-profile, env-audit | Audit of current project environment |
| Undo | undo-snapshot, rollback, revert | Roll back the last audit action |
Synonyms (audit = check = diagnosis, capability = skill = tool, environment = project) are automatically normalized into the corresponding mode.
Trigger Behavior
| Will Trigger ✅ | Will NOT Trigger ❌ |
|---|---|
skills-audit (sent standalone) | run skills-audit on this (in sentence) |
skill-check (sent standalone) | please do a skill-check (in sentence) |
project-audit (sent standalone) | I need a project-audit (in sentence) |
undo-snapshot (sent standalone) | undo that snapshot (in sentence) |
skills-audit deep (standalone + subcommand) | deep audit the skills (in sentence) |
Language Adaptation: If the user's primary language is English (>70% English in conversation), English trigger words are matched first, and the report output language switches to en.
When to Use
- Skill Audit Mode: Periodically maintain skill library health (recommended every 2-4 weeks) or after major changes to the skill library. Can be linked with
summarize: run summarize → discover skill accumulation → suggestskills-audit. - Project Audit Mode: Quickly understand the tech stack and tool chain when taking over a new project, or assess environment health before refactoring.
Prerequisites
Hard Requirements (auto-detected + install guidance if missing)
- codegraph — Code structure and symbol query, core skill analysis engine. If missing, attempts automatic installation.
- git — Version comparison, snapshots, and rollback. If missing, attempts automatic installation.
- config.yaml — The only configuration source in this skill's directory. If missing, generates from template.
Soft Requirements (automatically created if missing)
- user-profile.md — User profile. If missing, guides creation through Q&A.
Capabilities
skills-summarize-audit provides the following core capabilities (v6.0.0 adds capability dimension matrix, quality signals, community feed, and ecosystem radar):
- Project tech stack profiling and project type identification
- Full skill library scan and incremental inventory
- Capability dimension matrix construction (~12 dimensions, v6.0.0)
- Quality signal reading (linked with summarize error ledger, v6.0.0)
- Multi-factor effective capacity analysis and capacity warnings
- 6 skill activity checks
- 7-dimensional S/A/B/C/D scoring with mandatory justification
- T3 activity verification and false archiving risk control
- Community feed search (capability gap-driven GitHub search, v6.0.0)
- Tool recommendations based on capability complementarity (replace/complement/supplement/unrelated, v6.0.0)
- Ecosystem radar section (conditional output, only shown when issues exist, v6.0.0)
- Per-skill deep reading and external source verification
- External signal caching and 3-layer recommendation engine
- 30-second summary report and trend comparison
- Snapshot backup and
--undorollback - CI/CD non-interactive mode and JSON gating output
- Trigger word translation and description quality check
Detailed capability table by execution phase:
| Phase | Capability | Description |
|---|---|---|
| ⓪ Configuration | Multi-platform adaptation | Automatic detection of Codex/Claude/Cursor/Coze platform |
| ① Profiling | Project tech stack scan | Automatic identification of 15+ project types (Web/Rust/Python/Go/Java...) |
| ② Inventory | Full skill library scan + capability mapping | Scans 4 paths + capability-dimensions.yaml mapping (v6.0.0) |
| ②-bis | Capacity analysis | Multi-factor effective capacity calculation (token + conflict + cognitive capacity) |
| ②-ter | Activity check | 6 checks to verify skill usability (agent reachability/MCP/command/trigger/data/usage evidence) |
| ④ Scoring | 7-dimension scoring + quality signals | 7-dimension scoring + reads summarize error ledger (v6.0.0) |
| ④-a | T3 activity verification | Verifies usage evidence/modification time/trigger records before archiving to reduce false archiving rate |
| ④-bis | Deep reading | S/A full deep read, B (profile-related) full read, B (unrelated) summary |
| ⑤a | External signals | GitHub⭐/npm📥 cached collection |
| ⑤-aa | Community feed | Capability gap-driven GitHub search (v6.0.0) |
| ⑤b | Capability matrix recommendation engine | Recommendations based on capability complementarity (replace/complement/supplement/unrelated, v6.0.0) |
| ⑥ | Report generation + Ecosystem radar | 30-second summary + capacity block + score table + conditional ecosystem radar section (v6.0.0) |
| ⑦-a | Snapshot backup | Auto-backup all editable skills before execution |
| ⑦-d | Operation rollback | --undo to revert any execution, keeps 7-day snapshots |
| CI | Non-interactive audit | JSON output + health/t3/actions 3 gates, integrates with GitHub Actions |
When NOT to Use
- Single file edits, real-time debugging, minimal projects (no skill dependencies), or when only installing a single skill
- Low-resource environments (audit consumes ~5K-15K tokens per run)
Execution Flow
After triggering, read
references/execution-flow.mdfor the complete flow, scoring table, capacity calculation, report templates, and CI rules. Main file retains only entry points and key rules.
⓪ Pre-configuration → ① Project profiling → ② Skill inventory + capacity collection → **②-ter Activity check**
→ ③ Load reference data → ④ Layered scoring
→ ④-bis Deep reading → ⑤ External signals + recommendations → ⑥ Generate report → ⑥-bis Audit verification → ⑦ Execution → ⑧ Persistence
Key Rules
- Profiling: Only scan the current directory. Extract pain points from data sources. Do NOT guess. Write results to
{project}/.agents/project-profile.md. - Layering: T1 (hits core activities) → keep; T2 (general) → keep; T3 (specialized) → suggest archiving.
- Scoring: 7 dimensions S/A/B/C/D. Each skill must have 1 justification sentence; no justification = unscored.
- Capacity: Calculate Agent's maximum capacity. Can install ≤1 → 🔴, 2-4 → 🟡, ≥5 → 🟢.
- External Signals: GitHub⭐ cached for 7 days, npm📥 cached for 30 days. Skip if network is unavailable and annotate. Do NOT fabricate.
- Recommendations: Each with confidence score (0-10) + ROI. If <7, explicitly mark as "Evaluating"; if <5, require data support to recommend.
- Execution: Only modify user-level skills (editable=true). System skills are read-only.
- Golden Rule: When developing/modifying/recommending skills, search and verify first, citing ≥2 independent sources.
- Activity Check: First audit performs full check. Subsequent audits only check newly installed skills.
deepsubcommand forces full check. Trigger single-skill biopsy on anomalies.
Output Example
After triggering skills-audit, the 30-second summary block is output first:
🟢 Skill Library Health: Good
🛠️ Key Actions: Update 1 outdated skill, Archive 2 T3 skills
💰 Estimated Savings: ~3200 tokens/run (reducing low-value skill injection)
Followed by a score table snippet (example):
| Skill | Overall | Fit | Value | Fresh | Community | ROI | Novelty | Contamination | Suggestion |
|---|---|---|---|---|---|---|---|---|---|
| codegraph | S | 5 | 5 | 4 | 4 | 5 | 5 | 5 | Keep, core tool |
| old-linter | C | 2 | 2 | 2 | 3 | 2 | 2 | 4 | Suggest archiving |
Full report format in references/report-template.md.
Project Type Identification Rules
Core inference rules (full mapping in references/project-types.yaml):
package.json→ Web/Node/Full StackCargo.toml→ Rust ·pyproject.toml→ Python ·go.mod→ Gopom.xml/build.gradle→ Java ·.csproj→ C# ·pubspec.yaml→ FlutterCMakeLists.txt→ C/C++ ·Gemfile→ Ruby- Multiple
SKILL.md+MEMORY.md→AI Agent Skill Development
Boundary Rules (Quick Reference)
- Trigger: Standalone word trigger, NOT in sentence
- Layering: T1/T2/T3 based solely on profile "core activities"
- Scoring: 5 levels S/A/B/C/D. Mandatory justification. No justification = unscored.
- Deep Analysis: Read full body. Gaps must be verified with tools.
- Report: Output only if there is content. No fallback for empty sections.
- Recommendations: Confidence score 0-10 + ROI. Do not recommend if <5; require data support.
- Confirm Execution: User confirmation required for operations. System skills are read-only.
- Golden Rule: Search and verify first, ≥2 sources.
Platform Configuration & CI/CD
Supported platforms: platforms/zcode.yaml · platforms/workbuddy.yaml · platforms/claude.yaml · platforms/codex.yaml · platforms/cursor.yaml · platforms/default.yaml.
CI Mode (skills-audit --ci or skills-summarize-audit --ci): Non-interactive, JSON output, skip deep reading, bounded by ci.max_tokens budget in config.yaml. Gating logic: health=🔴 → exit_code=1 · actions_required>5 → exit_code=2 · t3_ratio>0.35 → exit_code=3. GitHub Actions template in references/ci-github-actions.yml.
SkillSpector Security Scan (optional): When enabled via config.yaml → security_scan, results are incorporated into the Value dimension (Critical→D, High→C, Pass→+0.5 grade). SkillSpector is an external process and does not consume Agent Tokens.
File Structure
skills-summarize-audit/
├── SKILL.md # Main skill file
├── VERSION # Version number
├── LICENSE # MIT
├── README.md # Overview and installation
├── config.yaml # Audit configuration
├── platforms/ # Platform adaptation configurations
├── references/ # Reference documents
│ ├── execution-flow.md # Execution flow index
│ ├── flow/ # Detailed steps per phase (19 sub-files)
│ ├── project-types.yaml # Project type mapping
│ ├── skill-registry.yaml # Skill registry
│ ├── deep-analysis-template.md
│ ├── description-quality.md
│ ├── report-template.md
│ ├── ci-output-schema.md
│ ├── ci-github-actions.yml
│ └── recommendation-examples.md
├── tests/ # Test suite and verification scripts
├── .data/ # Runtime data (logs/cache/statistics/profiles/external signals)
├── install.ps1 / install.sh # Installation scripts
└── CHANGELOG.md # Changelog
Installation
One-click installation and manual installation details in references/installation.md (if missing, refer to README.md).
After installation, enter skills-audit, skill-check, or skills-summarize-audit to trigger.
Feedback
License
MIT — see LICENSE file.