Understanding cangjie-skill, nuwa-skill, and darwin-skill: A Three-Layer AI Skill Ecosystem

cangjie-skill, nuwa-skill, and darwin-skill form a three-layer "skill ecosystem" where nuwa-skill distills human personas, cangjie-skill converts books into executable skills, and darwin-skill continuously evolves any skill through automated testing.

The kangarooking/cangjie-skill repository implements a sophisticated skill distillation framework that integrates with nuwa-skill and darwin-skill to create a complete pipeline for AI agent capabilities. Together, these three tools establish a layered architecture that transforms raw knowledge—whether from people or long-form content—into structured, testable, and self-improving Claude-compatible skills that agents can execute in real tasks.

The Three-Layer Skill Ecosystem

Each tool in this ecosystem targets a distinct knowledge source and provides a specialized execution model.

nuwa-skill: Distilling Human Personas

nuwa-skill focuses on people—their thinking style, expression DNA, and personal experience. Its core purpose is to create "human skills" that let AI agents mimic specific individuals (e.g., Elon Musk, Warren Buffett). The tool ingests persona-centric text such as tweets and interviews, extracts characteristic patterns, and wraps them in a Claude-compatible SKILL.md. According to the ecosystem description in README.en.md (lines 7-9), nuwa-skill represents the first layer of knowledge distillation.

cangjie-skill: Converting Books to Atomic Skills

cangjie-skill targets books and long-form content—methods, frameworks, principles, and case studies. It implements the RIA-TV++ pipeline (Adler analysis → parallel extraction → triple verification → RIA++ construction → Zettelkasten linking → pressure testing → delivery) to turn textual content into a set of atomic, executable skills. As specified in SKILL.md (lines 18-31), the pipeline produces a structured skill set including BOOK_OVERVIEW.md, INDEX.md, GLOSSARY.md, and individual SKILL.md files. Each methodological unit follows a six-section structure (R/I/A1/A2/E/B) that exposes reusable methodological units to agents.

darwin-skill: Automated Skill Evolution

darwin-skill operates on any existing skill, including those produced by nuwa-skill and cangjie-skill. It provides automatic evolution by continuously improving skills through feedback loops that evaluate performance and iteratively update skill definitions. The relationship is documented in README.en.md (lines 24-30), where darwin-skill is described as the "companion" that evolves any skill by consuming test data generated by the other tools.

Architectural Interaction

The three skills interlock through a standardized workflow that moves from raw data to refined, evolving capabilities.

Data Origin to Distillation

The pipeline begins with different input types for each tool. nuwa-skill receives person-centric text such as tweets and interviews, while cangjie-skill processes book-centric content including PDFs, EPUBs, and transcripts. This separation ensures that human behavioral patterns and structured methodologies are extracted using domain-appropriate techniques.

Distillation to Skill Definition

Both tools produce a SKILL.md following Claude's skill schema, but their internal structures reflect their distinct purposes. nuwa-skill emphasizes style, tone, and decision-making heuristics, whereas cangjie-skill follows the RIA-TV++ model with six specific sections for each methodological unit (R/I/A1/A2/E/B) as detailed in SKILL.md (lines 18-31). This structural difference ensures that book-derived skills expose actionable methodologies while persona-derived skills capture behavioral authenticity.

Testing and Evolution Integration

The ecosystem connection becomes concrete during the testing phase. cangjie-skill automatically generates test-prompts.json as part of its pressure-test stage (lines 31-35 of SKILL.md), including bait tests and cross-skill confusion tests. darwin-skill consumes these test prompts, runs them against the skill, and rewrites the skill's prompts, boundaries, and execution steps when performance drops below defined thresholds. This creates a closed loop where skills become increasingly robust through automated iteration.

Practical Implementation Workflow

Below are conceptual CLI examples illustrating how each skill operates within the ecosystem:

# 1️⃣ Distill a person (nuwa-skill)

nuwa-skill --source https://github.com/alchaincyf/nuwa-skill \
           --profile "elon_musk_tweets.txt" \
           --output ./skills/elon-musk

# 2️⃣ Distill a book (cangjie-skill)

cangjie-skill --text ./books/poor_charlies_almanack.txt \
              --title "Poor Charlie's Almanack" \
              --author "Charlie Munger" \
              --output ./skills/poor-charlies-almanack

# 3️⃣ Evolve a skill (darwin-skill)

darwin-skill --skill-dir ./skills/poor-charlies-almanack \
             --test-cases ./skills/poor-charlies-almanack/test-prompts.json \
             --iterations 5

These commands demonstrate the distinct roles: nuwa-skill for persona extraction, cangjie-skill for book distillation, and darwin-skill for continuous improvement.

Key Files Defining the Relationships

Several files within the cangjie-skill repository explicitly define these architectural relationships:

• SKILL.md (lines 54-60): Defines the meta-skill specification and explicitly positions cangjie-skill within the broader ecosystem alongside nuwa-skill and darwin-skill.
• README.en.md (lines 7-9, 24-30): Provides the high-level ecosystem narrative and comparative descriptions of all three tools.
• methodology/00-overview.md: Outlines the seven-stage RIA-TV++ pipeline that distinguishes cangjie-skill's processing depth from nuwa-skill's persona extraction.
• templates/SKILL.md.template: Contains the concrete R/I/A1/A2/E/B structure template used for atomic skill generation.

Summary

• cangjie-skill, nuwa-skill, and darwin-skill form a complete distill → structure → test → evolve pipeline for AI capabilities.
• nuwa-skill handles people-centric knowledge (personas, behaviors), while cangjie-skill handles book-centric knowledge (methods, frameworks) through the RIA-TV++ pipeline.
• darwin-skill consumes test-prompts.json generated by cangjie-skill (lines 31-35) to provide continuous automated improvement.
• The ecosystem enables AI agents to move beyond simple information retrieval to reliable, maintainable execution of complex skills.

Frequently Asked Questions

What is the main difference between nuwa-skill and cangjie-skill?

nuwa-skill distills human personas from sources like tweets and interviews to mimic specific individuals, while cangjie-skill distills books and long-form content into atomic, executable methodologies using the RIA-TV++ pipeline. nuwa-skill focuses on behavioral authenticity (style, tone, heuristics), whereas cangjie-skill focuses on structural methodology (R/I/A1/A2/E/B sections) derived from written works.

How does darwin-skill improve skills generated by cangjie-skill?

darwin-skill consumes the test-prompts.json file that cangjie-skill automatically generates during its pressure-testing stage (as implemented in SKILL.md lines 31-35). It runs these test cases against the skill, evaluates performance metrics, and iteratively rewrites the skill's prompts, boundaries, and execution steps when accuracy drops below defined thresholds, creating a continuous improvement loop.

Can cangjie-skill operate independently of nuwa-skill and darwin-skill?

Yes, cangjie-skill functions as a standalone tool for converting books into Claude-compatible skills. However, the full ecosystem value emerges when combining all three: nuwa-skill for persona capture, cangjie-skill for methodological extraction, and darwin-skill for ongoing evolution. The README.en.md (lines 7-9) explicitly positions them as complementary components of a unified skill architecture.

What does the RIA-TV++ pipeline in cangjie-skill entail?

The RIA-TV++ pipeline is a seven-stage process implemented in cangjie-skill: Adler analysis, parallel extraction, triple verification, RIA++ construction, Zettelkasten linking, pressure testing, and delivery. This methodology, detailed in SKILL.md (lines 18-31) and methodology/00-overview.md, ensures that book content is transformed into rigorous, testable, and interlinked atomic skills rather than simple summaries.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}