Skill Subtraction

Skill Subtraction (技能减法)

Audit your installed AI skills and cut the fat — a systematic, score-based review of every installed skill with clear keep / archive / uninstall recommendations.

Why subtraction (核心理念)

Most people keep adding skills — install one, see another, install that too — until dozens pile up and few get real use. Regular subtraction keeps the set lean:

• 认知清爽：技能越少，选择成本越低
• 资源聚焦：把精力投入到最有价值的技能上
• 维护省心：技能需要更新调试，越少负担越轻

Requirements (运行要求)

| Dependency | Requirement | Notes | |------|---------|---------| | Python | 3.10+ | Stdlib only, no third-party deps | | Runtime | python3 on PATH | The audit script is invoked by this skill | | Privileges | Non-root | Scan is read-only; uninstall/archive requires user confirmation | | Platforms | WorkBuddy / Codex / Claude Code / Cursor / Cline / Continue / LobsterAI | Follows the ~/./skills/ directory convention | | Env vars | None | No environment variables required |

Language auto-detection (语言自动检测)

Never ask the user to select a language upfront. Automatically detect and choose the report language based on the user's input:

• Chinese input / conversation → Output Chinese report directly, run script with --lang zh
• English input / conversation → Output English report directly, run script with --lang en
• Ambiguous / Undetectable input → Only if the language is truly ambiguous (e.g., pure numbers or code only), ask the user: "中文报告还是英文报告？ / Output in Chinese or English?"

Once determined, stick to that language for all workflow steps (scan, evaluation, report, confirmation).

Workflow (工作流程)

Step 1: Scan installed skills

Run the audit script; it auto-detects the hosting agent platform from its own path and scans that platform's installed skills (plus project-level skills in the current workspace). Pass --lang zh|en to match the conversation language:

python3 scripts/audit_skills.py --lang zh   # or --lang en
python3 scripts/audit_skills.py --agent codex
python3 scripts/audit_skills.py --all
python3 scripts/audit_skills.py --skills-dir "C:\Users\admin\AppData\Roaming\LobsterAI\SKILLs"
python3 scripts/audit_skills.py --workspace /path/to/workspace

Cross-platform notes: the script handles Windows GBK encoding and non-standard AppData/Roaming//SKILLs paths. Every failure point logs an issue into the JSON issues field and prints a stderr summary. Issue types: missing_skill_md, unreadable_skill_md, permission_denied, broken_symlink (error level); no_frontmatter, malformed_frontmatter, no_name_field, empty_description, not_a_directory (warning level). Exit codes: 0 = clean, 2 = scan done with error-level issues (CI-friendly).

Output is a JSON array; each entry includes name, agent, scope, path, description, agent_created, has_scripts, has_references, file_count, dir_size, last_modified, version. Plus source_stats (install-source counts) and batch_installs (≥ 5 skills created the same day → flagged as a batch).

Step 2: Classify & score

Apply the classification and scoring from the [Evaluation framework](#evaluation-framework-评估框架) section (full detail in references/evaluation_framework.md — read it for complex scenarios):

• 6 Functional Domains & Subcategories: Dev & System / Data & Connectors / Content & Media / Domain & Business / Productivity & Workflow / Meta & Agent Control
• Install Source (decoupled dimension): user-installed / platform-preinstalled / agent-created
• 6 weighted metrics (composite 24–100): usage frequency 25, necessity 20, current relevance 20, enabled status 15, maintenance 10, unique value 10

Step 3: Recommend

Map the composite score to keep / archive / uninstall using the decision matrix and special rules in the [Evaluation framework](#evaluation-framework-评估框架) section.

Step 4: Output the report

Using the language determined in the auto-detection section, output the matching template:

# 技能减法审计报告

**审计时间**：YYYY-MM-DD
**技能总数**：N 个（用户级 X 个，项目级 Y 个）

## 保留（N 个）

| 技能 | 类型 | 细分领域 | 保留理由 | 使用频率 |
|------|------|---------|---------|---------|
| ... | ... | ... | ... | ... |

## 归档（N 个）

| 技能 | 类型 | 细分领域 | 归档理由 | 重新激活条件 |
|------|------|---------|---------|------------|
| ... | ... | ... | ... | ... |

## 卸载（N 个）

| 技能 | 类型 | 细分领域 | 卸载理由 | 风险评估 |
|------|------|---------|---------|---------|
| ... | ... | ... | ... | ... |

## 汇总建议

- 当前技能集健康度：高/中/低
- 主要问题：...
- 下次审计建议时间：...

# Skill Subtraction Audit Report

**Audit Date**: YYYY-MM-DD
**Total Skills**: N (User-level: X, Project-level: Y)

## Keep (N)

| Skill | Type | Subcategory | Reason to Keep | Usage Frequency |
|-------|------|-------------|---------------|-----------------|
| ... | ... | ... | ... | ... |

## Archive (N)

| Skill | Type | Subcategory | Reason to Archive | Reactivation Condition |
|-------|------|-------------|--------------------|-----------------------|
| ... | ... | ... | ... | ... |

## Uninstall (N)

| Skill | Type | Subcategory | Reason to Uninstall | Risk Assessment |
|-------|------|-------------|---------------------|-----------------|
| ... | ... | ... | ... | ... |

## Summary

- Current skill set health: High/Medium/Low
- Main issues: ...
- Recommended next audit: ...

Step 5: Execute cleanup (user confirmation required)

After outputting the report, ask the user whether to execute cleanup. Never uninstall skills without consent.

• Execute cleanup: uninstall (via SkillManage) / archive (save SKILL.md and key config files to ~/./skill-archive/.md, then uninstall) / keep (no action)
• Report only: no action, the user decides later

Show each action before executing; proceed only after explicit confirmation.

Audit cycle recommendations (审计周期建议)

| Frequency | Scenario | |------|---------| | Quarterly | When skill count exceeds 10 | | After each project ends | Clean up project-level skills | | When business direction shifts | Re-evaluate business-type skills | | When feeling "too many skills" | Anytime |

Bundled resources (捆绑资源)

• scripts/audit_skills.py — auto-detects the hosting agent platform, scans all installed skills, parses frontmatter, outputs structured JSON
• references/evaluation_framework.md — full bilingual framework: classification, 6-metric scoring detail, decision matrix, special rules, dedup priority, archive standard and template. Read it for complex scenarios (batch dedup, archive recovery, platform-preinstalled batch filtering)
• examples/ — sample audit reports (English & Chinese), useful as expected-output references and demo material

Evaluation framework (评估框架)

> Core scoring tables and decision matrix below for daily use. Full detail (classification, dedup priority, archive standard & template) in references/evaluation_framework.md — read it for complex scenarios.

Six-metric scoring detail (六指标评分细则)

| Metric | Weight | Levels & scores | |------|------|---------| | Usage frequency | 25 | high 25 / medium 16 / low 8 / zero 4 | | Necessity | 20 | irreplaceable 20 / has alternatives 12 / nice-to-have 4 | | Current relevance | 20 | match 20 / partial 12 / irrelevant 4 | | Enabled status | 15 | enabled 15 / disabled but recently invoked 9 / disabled & never invoked 3 | | Maintenance | 10 | active ≤ 30d 10 / normal 30–90d 6 / stagnant > 90d 3 | | Unique value | 10 | unique 10 / partially unique 6 / complete overlap 3 |

Composite = weighted sum, range 24–100.

Decision matrix (判定矩阵)

| Score | Recommendation | Description | |------|------|---------| | 80–100 | Keep | High-value, master deeply | | 50–79 | Archive | Save config, uninstall, re-activate when needed | | 24–49 | Uninstall | Low value, clean up directly |

Special rules (特殊规则，覆盖评分)

1. Zero usage + irrelevant → uninstall
2. Complete overlap → keep the best one (dedup)
3. Disabled & never invoked → uninstall
4. Project-level + project ended → uninstall
5. Data source defunct → uninstall
6. Platform-preinstalled + batch + no match → batch archive (don't score individually, filter by business direction)
7. Platform-preinstalled + never triggered → archive (not uninstall; may be platform-dependent)
8. Batch detection: ≥ 5 skills created the same day (±1 day) → flag and evaluate as a group

Dedup & archive details (去重与归档细则)

• Dedup: overlapping skills keep only the best one; eliminated ones are marked "uninstall" with "overlaps with X" as the reason. Keep priority: complete → recently updated → high frequency → agent-created → lightweight (see references/evaluation_framework.md §4)
• Archive: target ~/./skill-archive/.md (see Step 5); full steps and file template in references/evaluation_framework.md §5

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: helloyxs
• Source: helloyxs/skill-subtraction
• License: MIT

Install and usage instructions live in the source repository linked above.