Skill FactorySAFE

oh-my-zsh for Claude Code — 16 agents, 35 commands, 32 skills, 21 safety hooks in one install. v4.0 adds an adversarial review loop: a second agent that never sees the first one's reasoning. MIT.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: skill-factory
description: >
  Analyze session work and automatically convert reusable patterns into Claude Code skills.
  Use when: "세션을 스킬로", "스킬 만들어", "이거 스킬로", "skill factory",
  "이 작업 자동화해", "스킬 추출", "make this a skill", "extract skill",
  "convert to skill", "스킬 팩토리", "자동 스킬 생성".
  Differs from skill-creator (archived) and manage-skills (drift detection):
  this skill actively analyzes sessions, checks for duplicates, and creates skills via Agent Teams.
disable-model-invocation: true
argument-hint: "[--dry-run] [--no-team] [--target name] [--scope global|project]"
---

# Skill Factory

Automated pipeline: session analysis -> duplicate check -> skill creation.
Requires: Python 3.8+, bash, git. Agent Teams path requires `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`.

| Existing Skill | Role | skill-factory Difference |
|----------------|------|--------------------------|
| skill-creator (archived) | Manual 6-step guide | Automated pipeline |
| manage-skills | Drift detection (verify-* skills) | Proactive skill generation (manage-skills verifies existing; skill-factory creates new) |
| continuous-learning | Passive pattern extraction | On-demand + team execution |

## Parameter Parsing

Parse `$ARGUMENTS` for flags:

| Flag | Default | Description |
|------|---------|-------------|
| `--dry-run` | false | Analyze and report only, no file creation |
| `--no-team` | false | Run sequentially without Agent Teams |
| `--target` | (auto) | Specific pattern name to extract |
| `--scope` | global | `global` (~/. claude/skills/) or `project` (.claude/skills/) |

If no arguments, run full auto-detection pipeline.

## Phase 1: Session Analysis

Collect what happened in this session:

```bash
# Uncommitted changes
git diff HEAD --name-only 2>/dev/null

# Recent commits on current branch
git log --oneline -20 2>/dev/null

# Branch diff from main
git diff main...HEAD --name-only 2>/dev/null
```

From collected changes, identify **candidate patterns** - repeatable workflows that appeared:

1. **Multi-step sequences** - 3+ actions performed in consistent order
2. **Tool combinations** - Specific tools used together (e.g., Grep + Read + Edit)
3. **Domain procedures** - File types or directories accessed with specific operations
4. **Repeated transformations** - Same type of change applied to multiple files

If `--target` is specified, focus analysis on that named pattern only.

For each candidate, produce a JSON entry (internal, not shown to user):

```json
{
  "name": "pattern-name",
  "description": "What was done repeatedly",
  "files": ["path/a.ts", "path/b.ts"],
  "steps": ["Step1", "Step2", "Step3"],
  "step_count": 3
}
```

Present findings to user:

```
Session Analysis Complete

Candidate Patterns Found: N

1. [pattern-name] - "Description of what was done repeatedly"
   Files: path/a.ts, path/b.ts (N files)
   Steps: Step1 -> Step2 -> Step3

2. [pattern-name] - "Description"
   ...

Which patterns should become skills? (select or 'all')
```

Wait for user selection before proceeding.

## Phase 2: Similarity Check

For each selected pattern, check against existing inventory.

**Step 1: Scan inventory**
```bash
bash $HOME/.claude/skills/skill-factory/scripts/scan-inventory.sh --scope all > /tmp/sf-manifest.json
```

**Step 2: Score similarity**
```bash
python3 $HOME/.claude/skills/skill-factory/scripts/similarity-scorer.py \
  --candidate "<pattern description>" \
  --candidate-name "<pattern-name>" \
  --manifest /tmp/sf-manifest.json \
  --top 3
```

**Step 3: Apply decision logic** (see [references/decision-tree.md](references/decision-tree.md))

Present results to user:

```
Similarity Check Results

Pattern: "pdf-batch-edit"
  Top match: nano-pdf (score: 0.72) -> MERGE
  Recommendation: Extend nano-pdf with batch operations

Pattern: "config-updater"
  Top match: init-project (score: 0.45) -> UPDATE
  Recommendation: Add config-update subsection to init-project

Pattern: "api-load-test"
  Top match: e2e (score: 0.24) -> CREATE
  Recommendation: Create new skill

Action for each pattern? (CREATE / UPDATE / MERGE / SKIP)
```

Wait for user decision per pattern.

## Phase 3: Blueprint

For each CREATE/UPDATE/MERGE decision, design the skill structure.

### CREATE Blueprint

Select template type from [references/skill-templates.md](references/skill-templates.md):
- **Workflow** for sequential processes
- **Task/Tool** for operation collections
- **Reference** for domain knowledge
- **Verification** for automated checks

Generate blueprint:

```
Blueprint: api-load-test

Type: Workflow
Scope: global (~/.claude/skills/)
Structure:
  api-load-test/
  ├── SKILL.md (~200 lines)
  │   ├── Frontmatter: name, description with triggers
  │   ├── Overview
  │   ├── Prerequisites
  │   ├── Workflow (4 steps)
  │   └── Output Format
  └── scripts/
      └── run-load-test.sh

Key sections:
  1. Target URL configuration
  2. Load profile definition
  3. Test execution
  4. Results analysis

Approve this blueprint? (y/n/edit)
```

Wait for user approval.

### UPDATE Blueprint

For UPDATE verdicts (score 0.3-0.6), plan a lightweight addition to the existing skill:

```
UPDATE Blueprint: config-updater -> init-project

Target skill: ~/.claude/skills/init-project/SKILL.md
Action: Add subsection "## Config Update" with steps
Estimated diff: +20-40 lines in existing SKILL.md
```

### MERGE Blueprint

For MERGE verdicts (score 0.6-0.8), plan a significant extension of the existing skill:

```
MERGE Blueprint: pdf-batch-edit -> nano-pdf

Target skill: ~/.claude/skills/nano-pdf/SKILL.md
Sections to add: "## Batch Operations" (new workflow section)
Scripts to add: scripts/batch-process.sh
Estimated diff: +60-100 lines in SKILL.md, +1 script
```

## Phase 4: Execution

Two paths based on `--no-team` flag and Agent Teams availability.

Check Agent Teams availability:
```bash
[ "${CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS:-0}" = "1" ] && echo "teams" || echo "no-team"
```
If `--no-team` is set or env var is missing/0, 

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.