meta-skill

Meta Skill

Create custom AI skills with guaranteed completeness and optimized retrieval. Meta-skill uses TDD + Anti-Rationalization Pressure Testing + Blind Comparison to ensure skill completeness, and redundancy removal + ambiguity clarification + progressive disclosure to maximize AI retrieval efficiency.

中文文档

---

Quick Start: Create Your First Skill

# In Qwen Code or Claude Code, simply ask:
"Create a skill for [your requirement]"

Example:

"Create a skill for automatic code review"
"Create a skill for writing unit tests"
"Create a skill for optimizing prompts"

Meta-skill will automatically:

Ensure Completeness:

1. TDD - Write tests first to define expected behavior
2. Anti-Rationalization Pressure Testing - Capture and plug loopholes under pressure scenarios
3. Blind Comparison - Verify candidate significantly outperforms baseline

Optimize AI Retrieval: 4. Ambiguity Clarification - Resolve unclear semantics 5. Redundancy Removal - Eliminate duplicate content 6. Progressive Disclosure - Structure information from simple to complex

7. Package as .skill file ready to use

---

Core Philosophy

Self-Evolution: The meta-skill uses its own pipeline to create and continuously improve skills (including itself) until convergence.

The skills/ directory contains the built-in skill library that meta-skill calls during its creation pipeline.

---

Core Flow

Intent Discovery → Type Decision → TDD Loop → Blind Comparison → AI Retrieval Optimization → Package

This README only keeps a lightweight flow view.

Single source of truth for the authoritative stage contract and gating rules:

• skills/meta-skill/SKILL.md

Stage (Lite View)	Main Components
Intent Discovery	intent-discovery
Type Decision	meta-skill stage-2 judgment (main type + enforcement tag)
TDD Loop	test-first + skill-format (+ anti-rationalization when enforcement tag is present)
Blind Comparison	agents/{grader,comparator,analyzer} + scripts/aggregate_benchmark.py
AI Retrieval Optimization	ai-doc-optimizer
Package	scripts/package_skill.py

---

Skill System Architecture

┌─────────────────────────────────────────────────────────────┐
│  skills/  (Built-in Skill Library)                          │
│                                                              │
│  ┌──────────────────────────────────────────────────────┐   │
│  │  meta-skill/ (Orchestrator)                          │   │
│  │  - SKILL.md                                          │   │
│  │  - agents/ (grader, analyzer, comparator)            │   │
│  │  - scripts/ (package_skill.py, aggregate_benchmark)  │   │
│  └──────────────────────────────────────────────────────┘   │
│                                                              │
│  ┌──────────────────────────────────────────────────────┐   │
│  │  Sub-skills (Called by meta-skill during pipeline)   │   │
│  │  - intent-discovery/  - test-first/                  │   │
│  │  - anti-rationalization/  - skill-format/            │   │
│  │  - ai-doc-optimizer/                                 │   │
│  └──────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Note: When creating a NEW skill, output goes to user-specified directory (~/.qwen/skills/, ./, etc.), NOT in meta-skill/skills/.

---

Skill Relationships

flowchart TB
    User[User Request] --> Meta[meta-skill<br/>Orchestrator]

    Meta --> ID[intent-discovery<br/>Requirement Clarification]
    ID --> Meta

    Meta --> TF[test-first<br/>TDD Methodology]
    TF --> AR[anti-rationalization<br/>Pressure Testing]
    AR --> TF
    TF --> SF[skill-format<br/>Format Validation]
    SF --> TF

    Meta --> AO[ai-doc-optimizer<br/>Iterative Optimization]
    AO --> AO

    subgraph Flow[Creation Flow]
        ID
        TF
        AO
    end

    subgraph Support[Support Skills]
        AR
        SF
    end

    Meta --> Flow
    Flow --> Support

---

Skills

Built-in Skill Library

These skills work together to create new skills:

Skill	Role in Skill Creation
meta-skill	Orchestrator — coordinates the entire skill creation pipeline
intent-discovery	Requirement Analyst — clarifies vague requirements through progressive questioning
test-first	TDD Engine — writes tests before implementation to ensure correctness
anti-rationalization	Quality Assurance — pressure-tests rules to prevent loopholes
skill-format	Validator — ensures SKILL.md follows proper format
ai-doc-optimizer	Optimizer — iteratively refines documentation for AI reading efficiency

How Skills Work Together

When you ask meta-skill to create a new skill: