skillforge（skillforge 分析）

SkillForge 4.1 - Intelligent Skill Router & Creator

Analyzes ANY input to find, improve, or create the right skill.

Quick Start

Any input works. SkillForge will intelligently route to the right action:

# These all work - SkillForge figures out what you need:

SkillForge: create a skill for automated code review
→ Creates new skill (after checking no duplicates exist)

help me debug this TypeError
→ Recommends ErrorExplainer skill (existing)

improve the testgen skill to handle React components better
→ Enters improvement mode for TestGen

do I have a skill for database migrations?
→ Recommends DBSchema, database-migration skills

TypeError: Cannot read property 'map' of undefined
→ Routes to debugging skills (error detected)

Triggers

Creation Triggers

• SkillForge: {goal} - Full autonomous skill creation
• create skill - Natural language activation
• design skill for {purpose} - Purpose-first creation
• ultimate skill - Emphasize maximum quality
• skillforge --plan-only - Generate specification without execution

Routing Triggers (NEW in v4.0)

• {any input} - Analyzes and routes automatically
• do I have a skill for - Searches existing skills
• which skill / what skill - Recommends matching skills
• improve {skill-name} skill - Enters improvement mode
• help me with / I need to - Detects task and routes

Process Overview

ANY USER INPUT
(prompt, error, code, URL, question, task request)
    │
    ▼
┌─────────────────────────────────────────────────────┐
│ Phase 0: SKILL TRIAGE (NEW)                         │
│ • Classify input type (create/improve/question/task)│
│ • Scan 250+ skills in ecosystem                     │
│ • Match against existing skills with confidence %   │
│ • Route to: USE | IMPROVE | CREATE | COMPOSE        │
├─────────────────────────────────────────────────────┤
│         ↓ USE_EXISTING    ↓ IMPROVE      ↓ CREATE   │
│      [Recommend]      [Load & Enhance] [Continue]   │
└─────────────────────────────────────────────────────┘
    │ (if CREATE_NEW or IMPROVE_EXISTING)
    ▼
┌─────────────────────────────────────────────────────┐
│ Phase 1: DEEP ANALYSIS                              │
│ • Expand requirements (explicit, implicit, unknown) │
│ • Apply 11 thinking models + Automation Lens        │
│ • Question until no new insights (3 empty rounds)   │
│ • Identify automation/script opportunities          │
├─────────────────────────────────────────────────────┤
│ Phase 2: SPECIFICATION                              │
│ • Generate XML spec with all decisions + WHY        │
│ • Include scripts section (if applicable)           │
│ • Validate timelessness score ≥ 7                   │
├─────────────────────────────────────────────────────┤
│ Phase 3: GENERATION                                 │
│ • Write SKILL.md with fresh context                 │
│ • Generate references/, assets/, and scripts/       │
├─────────────────────────────────────────────────────┤
│ Phase 4: SYNTHESIS PANEL                            │
│ • 3-4 Opus agents review independently              │
│ • Script Agent added when scripts present           │
│ • All agents must approve (unanimous)               │
│ • If rejected → loop back with feedback             │
└─────────────────────────────────────────────────────┘
    │
    ▼
Production-Ready Agentic Skill

Key principles:

• Phase 0 prevents duplicates - Always checks existing skills first
• Evolution/timelessness is the core lens (score ≥ 7 required)
• Every decision includes WHY
• Zero tolerance for errors
• Autonomous execution at maximum depth
• Scripts enable self-verification and agentic operation

Tool Escalation Policy

Start with least privilege (Read, Glob, Grep, Write, Edit).

Only add higher-risk tools when explicitly required:

• Bash for deterministic local scripts that cannot be replaced with file edits
• WebFetch / WebSearch only when external facts are required
• Task only for true parallel sub-agent orchestration

Commands

Phase 0: Skill Triage (NEW in v4.0)

Before creating anything, SkillForge intelligently analyzes your input to determine the best action.

How It Works

┌────────────────────────────────────────────────────────────────────┐
│                        ANY USER INPUT                               │
│  (prompt, error, code, URL, question, task request, anything)      │
└────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌────────────────────────────────────────────────────────────────────┐
│  Step 1: INPUT CLASSIFICATION                                       │
│  ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐       │
│  │ explicit_create │ │ explicit_improve│ │ skill_question  │       │
│  │ "create skill"  │ │ "improve skill" │ │ "do I have..."  │       │
│  └─────────────────┘ └─────────────────┘ └─────────────────┘       │
│  ┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐       │
│  │  task_request   │ │  error_message  │ │  code_snippet   │       │
│  │ "help me with"  │ │ "TypeError..."  │ │ [pasted code]   │       │
│  └─────────────────┘ └─────────────────┘ └─────────────────┘       │
└────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌────────────────────────────────────────────────────────────────────┐
│  Step 2: SKILL ECOSYSTEM SCAN                                       │
│  • Load index of 250+ skills (discover_skills.py)                  │
│  • Match input against all skills with confidence scoring          │
│  • Identify top matches with reasons                               │
└────────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌────────────────────────────────────────────────────────────────────┐
│  Step 3: DECISION MATRIX                                            │
│                                                                     │
│  Match ≥80%  + explicit create → CLARIFY (duplicate warning)       │
│  Match ≥80%  + other input     → USE_EXISTING (recommend skill)    │
│  Match 50-79%                  → IMPROVE_EXISTING (enhance match)  │
│  Match <50%  + explicit create → CREATE_NEW (proceed to Phase 1)   │
│  Multi-domain detected         → COMPOSE (suggest skill chain)     │
│  Ambiguous input               → CLARIFY (ask for more info)       │
└────────────────────────────────────────────────────────────────────┘

Decision Actions

Triage Script

# Run triage on any input
python scripts/triage_skill_request.py "help me debug this error"

# JSON output for automation
python scripts/triage_skill_request.py "create a skill for payments" --json

# Examples:
python scripts/triage_skill_request.py "TypeError: Cannot read property 'map'"
# → USE_EXISTING: Recommends ErrorExplainer (92%)

python scripts/triage_skill_request.py "create a skill for code review"
# → CLARIFY: CodeReview skill exists (85%), create anyway?

python scripts/triage_skill_request.py "help me with API and auth and testing"
# → COMPOSE: Multi-domain, suggests APIDesign + AuthSystem + TestGen chain

Ecosystem Index

Phase 0 uses a pre-built index of all skills:

# Rebuild skill index (run periodically or after installing new skills)
python scripts/discover_skills.py

# Index location: ~/.cache/skillrecommender/skill_index.json
# Scans: ~/.claude/skills/, plugins/marketplaces/*, plugins/cache/*

Integration with Phases 1-4

• USE_EXISTING: Exits early, no creation needed
• IMPROVE_EXISTING: Loads existing skill → Phase 1 analyzes gaps → Phase 2-4 enhance
• CREATE_NEW: Full pipeline (Phase 1 → 2 → 3 → 4)
• COMPOSE: Suggests using SkillComposer instead
• CLARIFY: Pauses for user input before proceeding

Validation & Packaging

Before distribution, validate your skill:

# Quick validation (required for packaging)
python scripts/quick_validate.py ~/.claude/skills/my-skill/

# Full structural validation
python scripts/validate-skill.py ~/.claude/skills/my-skill/

# Package for distribution
python scripts/package_skill.py ~/.claude/skills/my-skill/ ./dist

Frontmatter Requirements

Skills must use only these allowed frontmatter properties:

Basic Example:

---
name: my-skill
description: What this skill does and when to use it
license: MIT
model: claude-opus-4-5-20251101
user-invocable: true
metadata:
  version: 1.0.0
  author: your-name
---

Advanced Example (with forked context and hooks):

---
name: isolated-analyzer
description: Runs analysis in isolated context with validation hooks
license: MIT
model: claude-opus-4-5-20251101
context: fork
agent: Explore
user-invocable: true
allowed-tools:
  - Read
  - Glob
  - Grep
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate.sh"
metadata:
  version: 1.0.0
---

Field Details:

Skill Output Structure

~/.claude/skills/{skill-name}/
├── SKILL.md                    # Main entry point (required)
├── references/                 # Deep documentation (optional)
│   ├── patterns.md
│   └── examples.md
├── assets/                     # Templates (optional)
│   └── templates/
└── scripts/                    # Automation scripts (optional)
    ├── validate.py             # Validation/verification
    ├── generate.py             # Artifact generation
    └── state.py                # State management

Scripts Directory

Scripts enable skills to be agentic - capable of autonomous operation with self-verification.

Script Requirements:

• Python 3.x with standard library only (graceful fallbacks for extras)
• Result dataclass pattern for structured returns
• Exit codes: 0=success, 1=failure, 10=validation failure, 11=verification failure
• Self-verification where applicable
• Documented in SKILL.md with usage examples

See: references/script-integration-framework.md

Hooks Integration

Skills can define lifecycle hooks for validation, logging, and safety:

---
name: secure-skill
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "./scripts/validate-input.sh"
  PostToolUse:
    - matcher: "Write"
      hooks:
        - type: command
          command: "./scripts/log-output.sh"
          once: true
  Stop:
    - hooks:
        - type: command
          command: "./scripts/cleanup.sh"
---

Hook Types:

Hook Configuration:

Read $TOOL_INPUT / $TOOL_OUTPUT inside hook scripts from environment variables. Do not interpolate untrusted tool payloads directly into shell command strings.

When to Use Hooks:

Example: Script Validation Hook

For skills with scripts, add input validation:

hooks:
  PreToolUse:
    - matcher: "Bash(python:scripts/*)"
      hooks:
        - type: command
          command: "python scripts/quick_validate.py . 2>/dev/null || true"
          once: true

Anti-Patterns

Verification Checklist

After creation:

• Frontmatter valid (only allowed properties)
• Name is hyphen-case, ≤64 chars
• Description ≤1024 chars, no < or >
• 3-5 trigger phrases defined
• Timelessness score ≥ 7
• python scripts/quick_validate.py passes
• python scripts/check_docs_safety.py passes

1A: Input Expansion

Transform user's goal into comprehensive requirements:

USER INPUT: "Create a skill for X"
                │
                ▼
┌─────────────────────────────────────────────────────────┐
│ EXPLICIT REQUIREMENTS                                    │
│ • What the user literally asked for                      │
│ • Direct functionality stated                            │
├─────────────────────────────────────────────────────────┤
│ IMPLICIT REQUIREMENTS                                    │
│ • What they probably expect but didn't say               │
│ • Standard quality expectations                          │
│ • Integration with existing patterns                     │
├─────────────────────────────────────────────────────────┤
│ UNKNOWN UNKNOWNS                                         │
│ • What they don't know they need                         │
│ • Expert-level considerations they'd miss                │
│ • Future needs they haven't anticipated                  │
├─────────────────────────────────────────────────────────┤
│ DOMAIN CONTEXT                                           │
│ • Related skills that exist                              │
│ • Patterns from similar skills                           │
│ • Lessons from skill failures                            │
└─────────────────────────────────────────────────────────┘

Check for overlap with existing skills:

ls ~/.claude/skills/
# Grep for similar triggers in existing SKILL.md files

1B: Multi-Lens Analysis

Apply all 11 thinking models systematically:

Minimum requirement: All 11 lenses scanned, at least 5 applied in depth.

See: references/multi-lens-framework.md

1C: Regression Questioning

Iterative self-questioning until no new insights emerge:

ROUND N:
│
├── "What am I missing?"
├── "What would an expert in {domain} add?"
├── "What would make this fail?"
├── "What will this look like in 2 years?"
├── "What's the weakest part of this design?"
└── "Which thinking model haven't I applied?"
    │
    └── New insights > 0?
        │
        ├── YES → Incorporate and loop
        └── NO → Check termination criteria

Termination Criteria:

• Three consecutive rounds produce no new insights
• All 11 thinking models have been applied
• At least 3 simulated expert perspectives considered
• Evolution/timelessness explicitly evaluated
• Automation opportunities identified

See: references/regression-questions.md

1D: Automation Analysis

Identify opportunities for scripts that enable agentic operation:

FOR EACH operation in the skill:
│
├── Is this operation repeatable?
│   └── YES → Consider generation script
│
├── Does this produce verifiable output?
│   └── YES → Consider validation script
│
├── Does this need state across sessions?
│   └── YES → Consider state management script
│
├── Does this involve external tools?
│   └── YES → Consider integration script
│
└── Can Claude verify success autonomously?
    └── NO → Add self-verification script

Automation Lens Questions:

Decision: Script vs No Script

See: references/script-integration-framework.md

Specification Structure

The specification captures all analysis insights in XML format:

<skill_specification>
  <metadata>
    <name>skill-name</name>
    <analysis_iterations>N</analysis_iterations>
    <timelessness_score>X/10</timelessness_score>
  </metadata>

  <context>
    <problem_statement>What + Why + Who</problem_statement>
    <existing_landscape>Related skills, distinctiveness</existing_landscape>
  </context>

  <requirements>
    <explicit>What user asked for</explicit>
    <implicit>Expected but unstated</implicit>
    <discovered>Found through analysis</discovered>
  </requirements>

  <architecture>
    <pattern>Selected pattern with WHY</pattern>
    <phases>Ordered phases with verification</phases>
    <decision_points>Branches and defaults</decision_points>
  </architecture>

  <scripts>
    <decision_summary>needs_scripts + rationale</decision_summary>
    <script_inventory>name, category, purpose, patterns</script_inventory>
    <agentic_capabilities>autonomous, self-verify, recovery</agentic_capabilities>
  </scripts>

  <evolution_analysis>
    <timelessness_score>X/10</timelessness_score>
    <extension_points>Where skill can grow</extension_points>
    <obsolescence_triggers>What would break it</obsolescence_triggers>
  </evolution_analysis>

  <anti_patterns>
    <pattern>What to avoid + WHY + alternative</pattern>
  </anti_patterns>

  <success_criteria>
    <criterion>Measurable + verification method</criterion>
  </success_criteria>
</skill_specification>

See: references/specification-template.md

Specification Validation

Before proceeding to Phase 3:

• All sections present with no placeholders
• Every decision includes WHY
• Timelessness score ≥ 7 with justification
• At least 2 extension points documented
• All requirements traceable to source
• Scripts section complete (if applicable)
• Agentic capabilities documented (if scripts present)

Context: Fresh, clean (no analysis artifacts polluting) Standard: Zero errors—every section verified before proceeding

Generation Order

1. Create directory structure
   mkdir -p ~/.claude/skills/{skill-name}/references
   mkdir -p ~/.claude/skills/{skill-name}/assets/templates
   mkdir -p ~/.claude/skills/{skill-name}/scripts  # if scripts needed

2. Write SKILL.md
   • Frontmatter (YAML - allowed properties only)
   • Title and brief intro
   • Quick Start section
   • Triggers (3-5 varied phrases)
   • Quick Reference table
   • How It Works overview
   • Commands
   • Scripts section (if applicable)
   • Validation section
   • Anti-Patterns
   • Verification criteria
   • Deep Dive sections (in <details> tags)

3. Generate reference documents (if needed)
   • Deep documentation for complex topics
   • Templates for generated artifacts
   • Checklists for validation

4. Create assets (if needed)
   • Templates for skill outputs

5. Create scripts (if needed)
   • Use script-template.py as base
   • Include Result dataclass pattern
   • Add self-verification
   • Document exit codes
   • Test before finalizing

Quality Checks During Generation

Panel: 3-4 Opus agents with distinct evaluative lenses Requirement: Unanimous approval (all agents) Fallback: Return to Phase 1 with feedback (max 5 iterations)

Panel Composition

Script Agent (Conditional)

The Script Agent is activated when the skill includes a scripts/ directory. Focus areas:

Script Agent Scoring:

Agent Evaluation

Each agent produces:

## [Agent] Review

### Verdict: APPROVED / CHANGES_REQUIRED

### Scores
| Criterion | Score (1-10) | Notes |
|-----------|--------------|-------|

### Strengths
1. [Specific with evidence]

### Issues (if CHANGES_REQUIRED)
| Issue | Severity | Required Change |
|-------|----------|-----------------|

### Recommendations
1. [Even if approved]

Consensus Protocol

IF all agents APPROVED (3/3 or 4/4):
    → Finalize skill
    → Run validate-skill.py
    → Update registry
    → Complete

ELSE:
    → Collect all issues (including script issues)
    → Return to Phase 1 with issues as input
    → Re-apply targeted questioning
    → Regenerate skill and scripts
    → Re-submit to panel

IF 5 iterations without consensus:
    → Flag for human review
    → Present all agent perspectives
    → User makes final decision

See: references/synthesis-protocol.md

Every skill is evaluated through the evolution lens:

Temporal Projection

Timelessness Scoring

Requirement: All skills must score ≥7.

Anti-Obsolescence Patterns

See: references/evolution-scoring.md

Select based on task complexity:

Selection Decision Tree

Is it a simple procedure?
├── Yes → Single-Phase
└── No → Does it produce artifacts?
    ├── Yes → Generator
    └── No → Does it verify/audit?
        ├── Yes → Checklist
        └── No → Are subtasks independent?
            ├── Yes → Multi-Agent Parallel
            └── No → Multi-Agent Sequential or Multi-Phase

SKILLCREATOR_CONFIG:
  mode: autonomous
  depth: maximum  # always
  core_lens: evolution_timelessness

  analysis:
    min_lens_depth: 5
    max_questioning_rounds: 7
    termination_empty_rounds: 3

  synthesis:
    panel_size: 3
    require_unanimous: true
    max_iterations: 5
    escalate_to_human: true

  evolution:
    min_timelessness_score: 7
    min_extension_points: 2
    require_temporal_projection: true

  model:
    primary: claude-opus-4-5-20251101
    subagents: claude-opus-4-5-20251101

References

• Regression Questions - Complete question bank (7 categories)
• Multi-Lens Framework - 11 thinking models guide
• Specification Template - XML spec structure
• Evolution Scoring - Timelessness evaluation
• Synthesis Protocol - Multi-agent panel details
• Script Integration Framework - When and how to create scripts
• Script Patterns Catalog - Standard Python patterns

Related Skills

Extension Points

1. Additional Lenses: Add new thinking models to references/multi-lens-framework.md
2. New Synthesis Agents: Extend panel beyond 4 agents for specific domains
3. Custom Patterns: Add architecture patterns to selection guide
4. Domain Templates: Add domain-specific templates to assets/templates/
5. Script Patterns: Add new patterns to references/script-patterns-catalog.md
6. Script Categories: Extend the 7 script categories for new use cases

Changelog

v4.1.0 (Current)

• Extended frontmatter support - Full support for model, context, agent, hooks, user-invocable
• Created scripts/_constants.py for shared validation constants
• Updated scripts/quick_validate.py with extended property validation
• Updated scripts/validate-skill.py with hooks and agent validation
• Fixed scripts/discover_skills.py to extract version from metadata.version
• Added Hooks Integration section to SKILL.md and script-integration-framework.md
• Added Forked Context documentation to synthesis-protocol.md
• Updated skill template with modern best practices
• Expanded frontmatter requirements table with 10 properties

v4.0.0

• Phase 0 Skill Triage - Intelligent routing before creation
• Universal input handling - any prompt works
• Skill ecosystem scanning with 250+ skills indexed
• Decision matrix: USE | IMPROVE | CREATE | COMPOSE
• Renamed from SkillCreator to SkillForge

v3.2.0

• Added Script Integration Framework for agentic skills
• Added 4th Script Agent to synthesis panel (conditional)
• Added Phase 1D: Automation Analysis
• Added Automation Lens questions to regression questioning
• Created references/script-integration-framework.md
• Created references/script-patterns-catalog.md
• Created assets/templates/script-template.py
• Updated skill-spec-template.xml with <scripts> section
• Updated validate-skill.py with script validation
• Skills can now include self-verifying Python scripts

v3.1.0

• Added progressive disclosure structure
• Fixed frontmatter for packaging compatibility
• Added validation & packaging section
• Deep dive sections now collapsible

v3.0.0

• Complete redesign as ultimate meta-skill
• Added regression questioning loop
• Added multi-lens analysis framework (11 models)
• Added evolution/timelessness core lens
• Added multi-agent synthesis panel

v2.0.0

• Pattern selection guide
• Quality standards checklist

v1.0.0

• Basic skill structure