skill-from-history

name	skill-from-history
description	Analyzes Claude Code conversation history, git commits, and codebase to identify recurring patterns and generate project-specific agents and skills. Invoke with "/skill-from-history:learn-gen" or when user asks to "create skills from history", "analyze patterns", "generate agents".
compression-anchors	["Output: .claude/agents/[name]/AGENT.md + .claude/skills/[name]/SKILL.md","9-step process: Gather→Agents→Skills→Conflict→Propose→Generate","Governance: Maturity(Draft→Accepted→Canonical), Epochs, Staleness, Golden Tasks","Bidirectional: Agent→Skill (skills:) and Skill→Agent (agents:)","Global: ~/.claude/global-skills/ + Quarantine + Secret Redaction"]

Skill from History Generator

A meta-skill that discovers recurring patterns from your project's history and generates reusable Claude Code agents and skills. Agents define roles/perspectives; skills define workflows/procedures. Agents are generated first, then skills can reference them.

TL;DR

Input: Claude Code history + Git commits + Codebase Output: .claude/agents/[name]/AGENT.md + .claude/skills/[name]/SKILL.md Process: 9 steps (Gather → Agents → Skills → Conflict → Propose → Generate)

Quick Actions:

• /skill-from-history:learn-gen - Generate agents + skills (recommended)
• /skill-from-history:learn-gen --agents - Agents only
• /skill-from-history:learn-gen --skills - Skills only

---

When to Use

Invoke this skill when:

• Starting work on a mature project to capture existing patterns
• After completing significant features to document patterns used
• Periodically to discover new patterns from recent work
• When team members want to share project-specific knowledge

Trigger phrases: "create skills from history", "analyze patterns", "generate skills", "suggest skills", "skill from history"

---

Analysis Sources

Source	Location	Extracted Information
Conversation History	~/.claude/projects/[encoded-path]/	User prompts, tool calls, rejections
Git Commits	git log	Commit patterns, file frequencies
Codebase	Source files	Directory structure, code patterns

---

Execution Process (9 Steps)

1. Identify Project Path - Find Claude Code history location
2. Gather Sources - Parse JSONL files, git history, codebase
3. Check Existing - Scan .claude/skills/ and .claude/agents/
4. Extract Agents - Detect role/perspective patterns (60%+ confidence, 3+ occurrences)
5. Extract Patterns - Score by Frequency(30%), Reusability(25%), Complexity(20%), Documentation(15%), Specificity(10%)
6. Extract Anti-patterns - Detect rejected patterns ("No", "Don't", corrections)
7. Conflict Detection - Check duplicates (≥80% = update, 50-79% = extend, <50% = new)
8. Present Proposals - Display with confidence scores and recommendations
9. Generate & Validate - Create files with YAML frontmatter validation

詳細: references/execution-details.md

---

Generated Output

Directory Structure

.claude/
├── agents/[name]/
│   └── AGENT.md
├── skills/[name]/
│   ├── SKILL.md
│   └── references/
│       └── patterns.md
└── commands/[skill-name].md  # Optional

Skill Format

• YAML frontmatter: name (max 64 chars), description (max 1024 chars)
• Required sections: When to Use, Core Pattern, Constraints
• Evidence linking with [E#] references

詳細: references/skill-template.md, references/agent-template.md

---

Governance (v3.0)

Constraint Maturity

Level	説明	強制力
Draft	提案段階	Warning のみ
Accepted	承認済み	Warning + CI Warning
Canonical	標準規則	Error + CI Block
Deprecated	非推奨	Info のみ

昇格には: Eval Pass + Staleness Check + Rationale 記録が必要。

Key Concepts

• Epochs: プロジェクトの時代区分（古いルールは適用されない）
• Staleness: 時間減衰 + 再確認 + 違反でスコア計算
• Override: ALLOW_CONSTRAINT: CONS-XXX + REASON で例外許可

詳細: references/lifecycle.md

Golden Tasks (Eval Harness)

制約の有効性を検証するテストスイート:

/skill-from-history:learn-check            # 実行
/skill-from-history:learn-check --baseline # ベースライン比較

詳細: references/golden-tasks.md

---

Configuration

Setting	Default
Minimum frequency	3+ occurrences
Minimum confidence	60%
Git history	Last 100 commits
Analysis scope	Active source files only

---

Troubleshooting

Issue	Solution
No patterns found	Check history exists at ~/.claude/projects/
Too many similar suggestions	Increase similarity threshold, focus categories
Large history files	Limit to recent sessions, use sampling

---

Commands

コマンド	説明
/skill-from-history:learn-status	現在のステータス表示
/skill-from-history:learn-init	Golden Tasks初期化
/skill-from-history:learn-gen	Agent/Skill生成
/skill-from-history:learn-check	評価実行
/skill-from-history:learn-promote	成熟度昇格

---

Key Reminders

• Agent output: .claude/agents/[name]/AGENT.md (generated first)
• Skill output: .claude/skills/[name]/SKILL.md (can reference agents)
• Bidirectional: Agent→Skill (skills:) and Skill→Agent (agents:)
• Workflow: Observe → Draft → Promote → Refine
• Promotion rule: Eval Pass + Staleness Check + Rationale required
• Override: ALLOW_CONSTRAINT: CONS-XXX with REASON
• Global Skills: ~/.claude/global-skills/ for cross-project sharing
• Quarantine: External skills start as Draft + external: true

---

References

• execution-details.md - Step-by-step process details
• lifecycle.md - Maturity, Epochs, Staleness, Override
• golden-tasks.md - Evaluation harness
• skill-template.md - SKILL.md format
• agent-template.md - AGENT.md format
• agent-detection.md - Detection algorithms
• anti-pattern-detection.md - Negative learning
• conflict-detection.md - Duplicate handling
• validation.md - Validation rules
• global-skills.md - Cross-Project Learning