refactor-skill-structure
について
このスキルは、長すぎるまたは構成の悪いSKILL.mdファイルをリファクタリングし、CIの行数制限を満たし、可読性を向上させます。コード例を別ファイルに抽出し、複雑な手順を分割し、段階的な情報開示のためにコンテンツを再編成します。スキルが500行を超える場合、コードブロックが大部分を占める場合、または複数の無関係な操作を含む手順がある場合に使用してください。
クイックインストール
Claude Code
推奨npx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/refactor-skill-structureこのコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします
ドキュメント
Refactor Skill Structure
Refactor SKILL.md exceeded 500-line limit or w/ structural problems. Extract code examples to references/EXAMPLES.md, split compound procs into focused sub-procs, add cross-refs for progressive disclosure, verify skill complete + valid after restructure.
Use When
- Skill > 500-line CI limit
- Single proc step has multi unrelated ops → should be separate
- Code blocks > 15 lines dominate → could extract
- Skill accumulated ad-hoc sections breaking standard 6-section
- After update pushed over limit
- Review flagged structural issues beyond content
In
- Required: Path to SKILL.md
- Optional: Target line count (default 80% of 500 = ~400)
- Optional: Create
references/EXAMPLES.md? (default yes if extractable) - Optional: Split into multi skills? (default no, prefer extract first)
Do
Step 1: Measure + ID Bloat
Read skill + create section line budget → ID bloat.
# Total line count
wc -l < skills/<skill-name>/SKILL.md
# Line count per section (approximate)
grep -n "^## \|^### " skills/<skill-name>/SKILL.md
Classify bloat:
- Extractable: Code blocks > 15 lines, full configs, multi-variant examples
- Splittable: Compound proc steps doing 2+ unrelated ops
- Trimable: Redundant explanations, verbose ctx
- Structural: Ad-hoc sections not in standard 6
→ Line budget showing oversized sections + bloat category. Largest = primary refactor targets.
If err: skill < 500 lines + no structural issues → skill not needed. Verify request justified.
Step 2: Extract Code → references/EXAMPLES.md
Move code blocks > 15 lines to references/EXAMPLES.md, leave brief inline (3-10 lines) in main.
-
Create dir:
mkdir -p skills/<skill-name>/references/ -
For each extractable block:
- Copy full block to
references/EXAMPLES.mdw/ descriptive heading - Replace block in SKILL.md w/ brief 3-5 line snippet
- Add cross-ref:
See [EXAMPLES.md](references/EXAMPLES.md#heading) for the complete configuration.
- Copy full block to
-
Structure
references/EXAMPLES.mdw/ clear headings:# Examples ## Example 1: Full Configuration Complete configuration file for [context]: \```yaml # ... full config here ... \``` ## Example 2: Multi-Variant Setup ### Variant A: Development \```yaml # ... dev config ... \``` ### Variant B: Production \```yaml # ... prod config ... \```
→ All blocks > 15 lines extracted. Main SKILL.md keeps brief inline. Cross-refs link to extracted. references/EXAMPLES.md well-organized.
If err: extracting doesn't reduce enough (still > 500) → Step 3 splitting. Few code blocks (natural-lang skill) → focus Steps 3 + 4.
Step 3: Split Compound → Focused Steps
ID proc steps doing multi unrelated ops + split.
Signs compound step:
- Title contains "and" ("Configure Database and Set Up Caching")
- Step has multi Expected/On failure blocks (or should)
- Step > 30 lines
- Could skip or do diff order from sub-parts
For each compound:
- ID distinct ops in step
- Create new
### Step N:for each - Renumber subsequent
- Each new step → own Expected + On failure
- Add transition ctx between new steps
→ Each proc step does one thing. No step > 30 lines. Step count may grow but each indep verifiable.
If err: splitting → too granular (20+ total) → group related micro-steps under single step w/ numbered sub. Sweet spot 5-12 steps.
Step 4: Add Cross-Refs
Ensure main SKILL.md maintains readability + discoverability after extract.
For each extraction:
- Inline snippet in SKILL.md self-sufficient for common case
- Cross-ref explains additional content available
- Use relative paths:
[EXAMPLES.md](references/EXAMPLES.md#section-anchor)
Patterns:
- After brief snippet:
See [EXAMPLES.md](references/EXAMPLES.md#full-configuration) for the complete configuration with all options. - For multi-variant:
See [EXAMPLES.md](references/EXAMPLES.md#variants) for development, staging, and production variants. - For extended troubleshooting:
See [EXAMPLES.md](references/EXAMPLES.md#troubleshooting) for additional error scenarios.
→ Every extraction has cross-ref. Reader follows main for common case, drills into refs for detail.
If err: cross-refs make text awkward → consolidate multi refs into single note at end of step: For extended examples including [X], [Y], and [Z], see [EXAMPLES.md](references/EXAMPLES.md).
Step 5: Verify Line Count
Re-measure SKILL.md after changes.
# Check main SKILL.md
lines=$(wc -l < skills/<skill-name>/SKILL.md)
[ "$lines" -le 500 ] && echo "SKILL.md: OK ($lines lines)" || echo "SKILL.md: STILL OVER ($lines lines)"
# Check references file if created
if [ -f skills/<skill-name>/references/EXAMPLES.md ]; then
ref_lines=$(wc -l < skills/<skill-name>/references/EXAMPLES.md)
echo "EXAMPLES.md: $ref_lines lines"
fi
# Total content
echo "Total content: $((lines + ${ref_lines:-0})) lines"
→ SKILL.md < 500. Ideal < 400 → room future growth. references/EXAMPLES.md no limit.
If err: still > 500 after extract + split → skill should decompose into 2 separate skills. Too much ground = scope creep. Use create-skill for second + update Related Skills cross-refs both.
Step 6: Validate All Sections
After refactor, verify skill has all required sections + frontmatter intact.
Run review-skill-format checklist:
- YAML frontmatter parses
- All 6 required sections (When to Use, Inputs, Procedure, Validation, Common Pitfalls, Related Skills)
- Every proc step has Expected + On failure
- No orphaned cross-refs (all links resolve)
# Quick section check
for section in "## When to Use" "## Inputs" "## Procedure" "## Common Pitfalls" "## Related Skills"; do
grep -q "$section" skills/<skill-name>/SKILL.md && echo "$section: OK" || echo "$section: MISSING"
done
grep -qE "## Validation( Checklist)?" skills/<skill-name>/SKILL.md && echo "Validation: OK" || echo "Validation: MISSING"
→ All sections present. No content accidentally deleted during extract. Cross-refs in SKILL.md resolve to actual headings in EXAMPLES.md.
If err: section accidentally removed → restore from git: git diff skills/<skill-name>/SKILL.md. Cross-refs broken → verify heading anchors in EXAMPLES.md match links in SKILL.md (GitHub anchor: lowercase, hyphens for spaces, strip punctuation).
Check
- SKILL.md line count ≤ 500
- All code blocks in SKILL.md ≤ 15 lines
- Extracted in
references/EXAMPLES.mdw/ descriptive headings - Every extraction has cross-ref in main SKILL.md
- No compound proc steps remain (each step one thing)
- All 6 required sections present after refactor
- Every proc step has Expected: + On failure:
- YAML frontmatter intact + parseable
- Cross-ref links resolve to actual headings in EXAMPLES.md
-
review-skill-formatvalidation passes
Traps
- Extract too aggressive: All code → refs makes main unreadable. Keep 3-10 line snippets inline for common case. Only extract > 15 lines or multi-variant.
- Broken anchors: GitHub markdown anchors case-sensitive some renderers. Lowercase headings in EXAMPLES.md, match exact in cross-refs. Test
grep -c "heading-text" references/EXAMPLES.md. - Lose Expected/On failure during split: Each new step gets own Expected + On failure. Easy to leave one w/o blocks after split.
- Too many tiny steps: Splitting → 5-12 steps. End up 15+ → split too aggressive. Merge related micro back to logical groups.
- Forget update EXAMPLES.md headings: Rename section → all cross-ref anchors in SKILL.md must update. Grep old anchor to catch all refs.
→
review-skill-format— run format validation after refactor → confirm compliantupdate-skill-content— content updates often trigger structural refactor when push over limitcreate-skill— reference canonical structure when deciding how to organize extractedevolve-skill— split into 2 separate skills → use evolution to create derivative
GitHub リポジトリ
関連スキル
content-collections
メタこのスキルは、Content Collections(Markdown/MDXファイルを型安全なデータコレクションに変換するTypeScriptファーストのツール)の本番環境でテストされた設定を提供します。Zodバリデーションによる型安全性を実現し、ブログ、ドキュメントサイト、コンテンツ重視のVite + Reactアプリケーション構築時にご利用ください。Viteプラグインの設定、MDXコンパイルから、デプロイ最適化、スキーマバリデーションまで、すべてを網羅しています。
polymarket
メタこのスキルは、開発者がPolymarket予測市場プラットフォームを活用したアプリケーション構築を可能にします。API統合による取引や市場データの取得に加え、WebSocketを介したリアルタイムデータストリーミングにより、ライブ取引や市場活動を監視できます。取引戦略の実装や、ライブ市場更新を処理するツールの作成にご利用ください。
creating-opencode-plugins
メタこのスキルは、開発者がコマンド、ファイル、LSP操作など25種類以上のイベントタイプにフックするOpenCodeプラグインを作成することを支援します。JavaScript/TypeScriptモジュール向けに、プラグイン構造、イベントAPI仕様、および実装パターンを提供します。カスタムイベント駆動ロジックでOpenCode AIアシスタントのライフサイクルをインターセプト、監視、または拡張する必要がある場合にご利用ください。
sglang
メタSGLangは、高性能なLLMサービングフレームワークであり、RadixAttentionプレフィックスキャッシュを活用したJSON、正規表現、エージェントワークフロー向けの高速で構造化された生成を特長とします。特にプレフィックスが繰り返されるタスクにおいて、大幅に高速な推論を実現し、複雑な構造化出力やマルチターン対話に最適です。制約付きデコードが必要な場合や、広範なプレフィックス共有を伴うアプリケーションを構築する場合は、vLLMなどの代替案ではなくSGLangを選択してください。
