Skill Extractor

Analyze completed task outcomes to identify reusable patterns and propose new SKILL.md candidates. Inspired by Hermes Agent's self-learning skill extraction — adapted for oh-my-customcode's compilation metaphor.

Philosophy

In the compilation metaphor: task trajectories are runtime traces, and extracted skills are new source code. This skill turns successful execution patterns into reusable knowledge artifacts.

Runtime traces (task outcomes) → Pattern analysis → SKILL.md proposal → User approval → mgr-creator

Usage

/skill-extractor                    # Analyze current session outcomes
/skill-extractor --threshold 2      # Lower success threshold (default: 3)
/skill-extractor --dry-run          # Preview proposals without writing

Options

--threshold, -t   Minimum success count for pattern qualification (default: 3)
--dry-run, -d     Preview proposals to stdout only, no file writes
--all             Include all sessions (not just current, requires task outcome history)

Workflow

Phase 1: Collect Task Outcomes

Read task outcome data from the session:

# Current session outcomes (from task-outcome-recorder hook)
OUTCOMES_FILE="/tmp/.claude-task-outcomes-${PPID}"

If file doesn't exist or is empty: report "No task outcomes recorded in this session." and stop.

Parse JSONL entries. Each entry has:

{"agent_type": "lang-typescript-expert", "skill": "typescript-best-practices", "description": "Fix type error in auth module", "outcome": "success", "model": "sonnet", "timestamp": "2026-04-05T09:30:00Z", "duration_ms": 15000}

Phase 2: Pattern Detection

Group outcomes by (agent_type, skill) tuple:

Pattern: (lang-typescript-expert, typescript-best-practices)
  → success: 5, failure: 1, total: 6
  → success_rate: 0.83
  → descriptions: ["Fix type error...", "Refactor module...", ...]

Filter qualifying patterns:

• success_count >= threshold (default: 3)
• success_rate >= 0.8
• Not already an existing skill (check .claude/skills/*/SKILL.md)

Phase 3: Generate Proposals

For each qualifying pattern, generate a SKILL.md proposal:

## Proposal: {proposed-skill-name}

**Source Pattern**: {agent_type} + {skill} ({success_count} successes, {success_rate}% rate)
**Confidence**: {low|medium|high} (based on count and rate)

### Proposed SKILL.md

name: {proposed-name}
description: {inferred from common description patterns}
scope: core
user-invocable: false

### Rationale
{Why this pattern should be extracted as a skill — based on frequency and success rate}

### Overlap Check
{List any existing skills with >50% keyword overlap}

Confidence scoring:

Successes	Rate	Confidence
3-5	>= 0.8	low
6-10	>= 0.85	medium
10+	>= 0.9	high

Phase 4: Present to User

Display proposals in ranked order (highest confidence first):

[skill-extractor] {N} skill candidates detected

  1. [high] proposed-skill-name
     Source: {agent_type} + {skill} (12 successes, 92%)
     Description: {inferred description}

  2. [medium] another-skill-name
     Source: {agent_type} + {skill} (7 successes, 86%)
     Description: {inferred description}

Select [1-N] to create, "all" to create all, or "skip" to cancel:

Phase 5: Create Skill (on approval)

Delegate to mgr-creator with the proposal context:

• Proposed name and description
• Source pattern data
• Confidence level
• Any overlap warnings

mgr-creator handles: SKILL.md creation, template sync, ontology registration.

Selection Discipline (evidence-gated)

Before proposing a SKILL candidate, apply this gate. Default to NOT creating a new skill — prefer strengthening an existing skill/rule.

Evidence Hierarchy

Rank supporting evidence; only direct, repeated success qualifies:

Tier	Evidence	Action
1 Direct	Pattern executed successfully ≥2 times in observed trajectories	Eligible to propose
2 Inferred	Pattern plausible but observed once	Hold — do not propose yet
3 Speculative	Pattern imagined from a single description	Reject

4-Criteria Selection Gate

A pattern becomes a candidate only if ALL four hold:

1. Reusable across ≥2 distinct contexts (not one-off)
2. Non-trivial — encodes real workflow knowledge, not a single command
3. Not already covered by an existing skill (run an overlap check first)
4. Demonstrated success ≥2 times (Evidence Hierarchy tier 1)

Two-Phase Restraint

• Phase 1 (broad): collect all repeated patterns as raw candidates.
• Phase 2 (restrain): filter through the 4-criteria gate. Default to NOT creating a skill; prefer strengthening an existing skill/rule over spawning a new one.

Borrowed from /scout #1268 (evidence-hierarchy + selection gate + two-phase restraint). Reference: issue #1268.

Integration

System	How
task-outcome-recorder	Reads JSONL outcomes as input data
feedback-collector	Complementary: feedback-collector extracts failure patterns, skill-extractor extracts success patterns
mgr-creator	Delegated skill creation on user approval
skills-sh-search	Check agentskills.io for existing equivalent before creating
R011 (memory)	User Model tracks extraction decisions in Override Decisions

Hook Integration

The skill-extractor-analyzer.sh Stop hook provides a lightweight pre-analysis:

• Reads task outcomes file
• Counts qualifying patterns
• Emits advisory stderr message if candidates found
• Does NOT create skills (that requires user approval via the skill)

Safety

• User approval required: Never auto-creates skills
• Overlap check: Prevents duplicating existing skills
• Dry-run mode: Preview without side effects
• Advisory hook: Stop hook is advisory-only (exit 0)
• Confidence transparency: All proposals show confidence scores

--mode failure (Skillify Pattern)

feedback memory에 누적된 실패 패턴을 분석하여 영구 구조(스킬 또는 규칙 확장)로 전환하는 모드.

입력

• .claude/agent-memory*/feedback_*.md (누적된 실패 메모리)
• MEMORY.md의 Feedback Memories 섹션

처리

1. 각 feedback memory의 Why/How to apply 필드에서 공통 패턴 추출
2. 3회 이상 반복되는 패턴을 "failure candidate"로 격상
3. 후보 각각에 대해:
   • 기존 스킬 확장으로 해결 가능? → 스킬 업데이트 제안
   • 규칙 명문화가 더 적합? → R016 Matrix "Skill Promotion" 열에 등록
   • 신규 스킬이 필요? → context fork cap (12/12) 여부 확인 후 제안

출력

.claude/outputs/sessions/{date}/skill-extractor-failure-{HH}.md 아티팩트 (R006 Artifact Channel Protocol)

Tool: Writing artifacts under .claude/outputs/

Under mode: "bypassPermissions", subagents write directly to .claude/outputs/sessions/ with the Write tool — direct .claude/** writes are permitted (CC v2.1.121+, #1101). No /tmp staging or script wrapping is needed. Read-only Bash on .claude/outputs/ (e.g., cat, head, wc) is allowed for verification.

Reference: R006/R010 sensitive-path handling (direct .claude/** write under bypassPermissions), #1101.

참조

• R016 MUST-continuous-improvement.md Defect Response Matrix — Skill Promotion 열
• Skillify 내재화 배경: issue #972
• context fork cap: .claude/rules/MUST-agent-design.md Skill Frontmatter "Context Fork Criteria"