evolve-skill
について
`evolve-skill`スキルは、既存のスキルの内容をその場で改良するか、より高度なバリアントを作成することで、スキルを更新します。要件の収集から変更の適用、バージョンメタデータの更新、相互参照の同期まで、プロセス全体を扱います。スキルの手順が古くなっている場合、フィードバックから不足が明らかになった場合、または複雑さの向上が必要な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/evolve-skill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
Evolve an Existing Skill
Improve, extend, or create an advanced variant of a skill that was originally authored with create-skill. This procedure covers the maintenance side of the skill lifecycle: assessing gaps, applying targeted improvements, bumping versions, and keeping the registry and cross-references in sync.
When to Use
- A skill's procedure steps are outdated or incomplete after tooling changes
- User feedback reveals missing pitfalls, unclear steps, or weak validation
- A skill needs to grow from basic to intermediate (or intermediate to advanced)
- An advanced variant is needed alongside the original (e.g.,
create-r-packageandcreate-r-package-advanced) - Related skills were added or removed and cross-references are stale
Inputs
- Required: Path to the existing SKILL.md to evolve
- Required: Evolution trigger (feedback, tooling change, complexity upgrade, new related skills, discovered pitfalls)
- Optional: Target complexity level if changing (basic, intermediate, advanced)
- Optional: Whether to create an advanced variant instead of refining in-place (default: refine in-place)
Procedure
Step 1: Assess the Current Skill
Read the existing SKILL.md and evaluate each section against the quality checklist:
| Section | What to Check | Common Issues |
|---|---|---|
| Frontmatter | All required fields present, description < 1024 chars | Missing tags, stale version |
| When to Use | 3-5 concrete trigger conditions | Vague or overlapping triggers |
| Inputs | Required vs optional clearly separated | Missing defaults for optional inputs |
| Procedure | Each step has code + Expected + On failure | Missing On failure blocks, pseudocode instead of real commands |
| Validation | Each item is binary pass/fail | Subjective criteria ("code is clean") |
| Common Pitfalls | 3-6 with cause and avoidance | Too generic ("be careful") |
| Related Skills | 2-5 valid skill references | Stale references to renamed/removed skills |
# Read the skill
cat skills/<skill-name>/SKILL.md
# Check frontmatter parses
head -20 skills/<skill-name>/SKILL.md
# Verify related skills still exist
grep -oP '`[\w-]+`' skills/<skill-name>/SKILL.md | sort -u
Got: A list of specific gaps, weaknesses, or improvement opportunities.
If fail: If the SKILL.md doesn't exist or has no frontmatter, this skill doesn't apply — use create-skill instead to author it from scratch.
Step 2: Gather Evolution Requirements
Identify and categorize what triggered the evolution:
| Trigger | Example | Typical Scope |
|---|---|---|
| User feedback | "Step 3 is unclear" | Refinement |
| Tooling change | New API version, deprecated command | Refinement |
| Discovered pitfall | Common failure not documented | Refinement |
| Complexity upgrade | Skill is too shallow for real use | Refinement or variant |
| New related skills | Adjacent skill was added | Refinement (cross-refs) |
| Advanced use case | Power users need deeper coverage | Variant |
Document the specific changes needed before editing. List each change with its target section.
Got: A concrete list of changes (e.g., "Add On failure to Step 4", "Add new Step 6 for edge case X", "Update Related Skills to include new-skill").
If fail: If the changes are unclear, consult the user for clarification before proceeding. Vague evolution goals produce vague improvements.
Step 3: Choose Evolution Scope
Use this decision matrix to determine whether to refine in-place or create a variant:
| Criteria | Refinement (in-place) | Advanced Variant (new skill) |
|---|---|---|
| Skill ID | Unchanged | New ID: <skill>-advanced |
| File path | Same SKILL.md | New directory |
| Version bump | Patch or minor | Starts at 1.0 |
| Complexity | May increase | Higher than original |
| Registry | No new entry | New entry added |
| Symlinks | No change | New symlinks needed |
| Original skill | Modified directly | Left intact, gains cross-reference |
Refinement: Choose when improving quality, fixing gaps, or adding modest new content. The skill keeps its identity.
Variant: Choose when the evolved version would double the length, change the target audience, or require substantially different inputs. The original stays as-is for simpler use cases.
Got: A clear decision — refinement or variant — with rationale.
If fail: If unsure, default to refinement. You can always extract a variant later; it's harder to merge one back.
Step 4: Apply Content Changes
For Refinements
Edit the existing SKILL.md directly:
# Open for editing
# Add/revise procedure steps
# Strengthen Expected/On failure pairs
# Add tables or examples
# Update When to Use triggers
# Revise Inputs if scope changed
Follow these editing rules:
- Preserve all existing sections — add content, don't remove sections
- Keep step numbering sequential after insertions
- Every new or modified step must have both Expected and On failure
- New pitfalls go at the end of the Common Pitfalls section
- New related skills go at the end of the Related Skills section
For Variants
# Create the variant directory
mkdir -p skills/<skill-name>-advanced/
# Copy the original as a starting point
cp skills/<skill-name>/SKILL.md skills/<skill-name>-advanced/SKILL.md
# Edit the variant:
# - Change `name` to `<skill-name>-advanced`
# - Update `description` to reflect the advanced scope
# - Raise `complexity` (e.g., intermediate → advanced)
# - Reset `version` to "1.0"
# - Add/expand procedure steps for the advanced use case
# - Reference the original in Related Skills as a prerequisite
Got: The SKILL.md (refined or new variant) passes the assessment checklist from Step 1.
If fail: If a step edit breaks the document structure, use git diff to review changes and revert partial edits with git checkout -- <file>.
Step 4.5: Sync Translated Variants
Required when translations exist. This step applies to both human authors and AI agents following this procedure. Do not skip — stale
source_commitvalues causenpm run validate:translationsto report false staleness warnings across all locales.
Check whether translations exist for the evolved skill and update them to reflect the new source state:
# Check for existing translations
ls i18n/*/skills/<skill-name>/SKILL.md 2>/dev/null
If translations exist
- Get the current source commit hash:
SOURCE_COMMIT=$(git rev-parse HEAD)
- Update
source_commitin each translated file's frontmatter:
for locale_file in i18n/*/skills/<skill-name>/SKILL.md; do
sed -i "s/^source_commit: .*/source_commit: $SOURCE_COMMIT/" "$locale_file"
done
- Flag files for re-translation by including affected locales in the commit message:
evolve(<skill-name>): <description of changes>
Translations flagged for re-sync: de, zh-CN, ja, es
Changed sections: <list sections that changed>
- Regenerate translation status files:
npm run translation:status
If no translations exist
No action needed. Proceed to Step 5.
For variants
Defer translation of new variants until the variant stabilizes (1-2 versions). Translating a v1.0 variant that may change substantially by v1.2 wastes effort. Add translations after the variant has been refined at least once.
Got: All translated files have source_commit updated to the current commit. The commit message notes which locales need re-translation and which sections changed. npm run translation:status exits 0.
If fail: If sed fails to match the frontmatter field, the translated file may have non-standard formatting. Open it manually and verify it has source_commit in its YAML frontmatter. If the field is missing, the file was not scaffolded correctly — re-scaffold with npm run translate:scaffold.
Step 5: Update Version and Metadata
Bump the version field in frontmatter following semver conventions:
| Change Type | Version Bump | Example |
|---|---|---|
| Typo fix, wording clarification | Patch: 1.0 → 1.1 | Fixed unclear sentence in Step 3 |
| New step, new pitfall, new table | Minor: 1.0 → 2.0 | Added Step 7 for edge case handling |
| Restructured procedure, changed inputs | Major: 1.0 → 2.0 | Reorganized from 5 to 8 steps |
Also update:
complexityif the scope expanded (e.g., basic → intermediate)tagsif the coverage area changeddescriptionif the skill's scope is materially different
Got: Frontmatter version reflects the magnitude of changes. New variants start at "1.0".
If fail: If you forget to bump the version, the next evolution will have no way to distinguish the current state from the previous one. Always bump before committing.
Step 6: Update Registry and Cross-References
For Refinements
No registry changes are needed (path unchanged). Update cross-references only if Related Skills changed in other skills:
# Check if any skill references the evolved skill
grep -r "<skill-name>" skills/*/SKILL.md
For Variants
Add the new skill to skills/_registry.yml:
- id: <skill-name>-advanced
path: <skill-name>-advanced/SKILL.md
complexity: advanced
language: multi
description: One-line description of the advanced variant
Then:
- Increment
total_skillsat the top of the registry - Add Related Skills cross-reference in the original skill pointing to the variant
- Add Related Skills cross-reference in the variant pointing to the original
- Create symlinks for slash command discovery:
# Project-level
ln -s ../../skills/<skill-name>-advanced .claude/skills/<skill-name>-advanced
# Global
ln -s /mnt/d/dev/p/agent-almanac/skills/<skill-name>-advanced ~/.claude/skills/<skill-name>-advanced
Got: Registry total_skills matches find skills -name SKILL.md | wc -l. Cross-references are bidirectional.
If fail: If the registry count is wrong, run find skills -name SKILL.md | wc -l to get the true count and correct the registry. For broken symlinks, use readlink -f to debug resolution.
Step 7: Validate the Evolved Skill
Run the full validation checklist:
- SKILL.md exists at the expected path
- YAML frontmatter parses without errors
-
versionwas bumped (refinement) or set to "1.0" (variant) - All sections present: When to Use, Inputs, Procedure, Validation, Common Pitfalls, Related Skills
- Every procedure step has Expected and On failure blocks
- Related Skills reference valid, existing skill names
- Registry entry exists with correct path (variants only)
-
total_skillscount matches actual skill count on disk - Symlinks resolve correctly (variants only)
-
git diffshows no accidental deletions from the original content - For refinements with translations:
source_commitupdated or translations flagged for re-sync
# Verify frontmatter
head -20 skills/<skill-name>/SKILL.md
# Count skills on disk vs registry
find skills -name SKILL.md | wc -l
grep total_skills skills/_registry.yml
# Check symlinks (for variants)
ls -la .claude/skills/<skill-name>-advanced
readlink -f .claude/skills/<skill-name>-advanced/SKILL.md
# Review all changes
git diff
Got: All checklist items pass. The evolved skill is ready to commit.
If fail: Address each failing item individually. The most common post-evolution issue is a stale total_skills count — always verify it last.
Validation
- SKILL.md exists and has valid YAML frontmatter
-
versionfield reflects the changes made - All procedure steps have Expected and On failure blocks
- Related Skills references are valid (no broken cross-references)
- Registry
total_skillsmatches actual count on disk - For variants: new entry in
_registry.ymlwith correct path - For variants: symlinks created at
.claude/skills/and~/.claude/skills/ -
git diffconfirms no accidental content removal - For refinements with translations:
source_commitupdated or translations flagged for re-sync
Pitfalls
- Forgetting to bump version: Without version bumps, there's no way to track what changed or when. Always update
versionin frontmatter before committing. - Accidental content deletion: When restructuring steps, it's easy to drop an On failure block or a table row. Always review
git diffbefore committing. - Stale cross-references: When creating a variant, both the original and the variant need to reference each other. One-directional references leave the graph incomplete.
- Registry count drift: After creating a variant, the
total_skillscount must be incremented. Forgetting this causes validation failures in other skills that check the registry. - Stale translations after evolution: With 1,288 translation files in the repo, every skill evolution triggers staleness in up to 4 locale files. Always check for existing translations with
ls i18n/*/skills/<skill-name>/SKILL.mdand updatesource_commitin each translated file's frontmatter, or flag them for re-translation in the commit message. Skipping this causesnpm run validate:translationsto report stale warnings. - Scope creep during refinement: A refinement that doubles the skill's length should probably be a variant instead. If you're adding more than 3 new procedure steps, reconsider the scope decision from Step 3.
- Avoid
git mvon NTFS-mounted paths (WSL): On/mnt/paths,git mvfor directories can create broken permissions (d?????????). Usemkdir -p+ copy files +git rmthe old path instead. See the environment guide troubleshooting section.
Related Skills
create-skill— foundation for authoring new skills; evolve-skill assumes this was followed originallycommit-changes— commit the evolved skill with a descriptive messageconfigure-git-repository— version-controlled skill changessecurity-audit-codebase— review evolved skills for accidentally included secrets
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。
