Cangjie-Skill vs Nuwa-Skill vs Darwin-Skill: A Three-Layer Knowledge Distillation Ecosystem

Cangjie-skill, nuwa-skill, and darwin-skill form an integrated three-layer ecosystem where nuwa-skill distills human personas, cangjie-skill converts books into executable atomic skills, and darwin-skill continuously evolves both through automated feedback loops.

The kangarooking/cangjie-skill repository serves as the book-distillation engine in a comprehensive AI skill architecture designed to transform static knowledge into agentic capabilities. When evaluating cangjie-skill vs nuwa-skill vs darwin-skill, you are examining a sequential pipeline that converts raw inputs—whether human expression, published works, or existing skill definitions—into tested, self-improving operational tools.

The Three-Layer Distillation Architecture

Nuwa-Skill: Capturing Human Expression DNA

Nuwa-skill targets people-centric knowledge, extracting thinking styles, expression patterns, and personal heuristics from textual outputs like tweets and interviews. According to the ecosystem description in README.en.md, it creates "human skills" that allow AI agents to mimic specific individuals by wrapping characteristic patterns in Claude-compatible SKILL.md files.

Cangjie-Skill: Structuring Book Knowledge

Cangjie-skill processes book-centric and long-form content—including PDFs, EPUBs, and transcripts—into atomic, executable skill sets. As implemented in SKILL.md, it runs the RIA-TV++ pipeline, a seven-stage process involving Adler analysis, parallel extraction, triple verification, RIA++ construction, Zettelkasten linking, pressure testing, and delivery. Each resulting skill follows a strict six-section structure (R/I/A1/A2/E/B) defined in templates/SKILL.md.template, producing deliverables like BOOK_OVERVIEW.md, INDEX.md, GLOSSARY.md, and individual SKILL.md files.

Darwin-Skill: Automated Skill Evolution

Darwin-skill consumes the output of both previous layers to provide continuous improvement of any existing skill. As noted in the README.en.md ecosystem section (lines 24-30), it operates as the "companion" evolution engine, ingesting test-prompts.json files generated during cangjie-skill's pressure-testing stage and iteratively rewriting skill definitions when performance metrics drop below defined thresholds.

Architectural Interaction and Execution Flow

The interaction between these components follows a strict data lineage:

1. Data Ingestion and Source Differentiation

• Nuwa-skill receives person-centric text streams (social media, interviews, essays).
• Cangjie-skill receives book-centric documents (scanned texts, transcripts, course materials).

2. Skill Definition Generation

Both tools output Claude-compatible SKILL.md files, but with distinct internal architectures:

• Nuwa-skill emphasizes tone, stylistic markers, and decision-making heuristics.
• Cangjie-skill implements the RIA-TV++ model, exposing six structured sections (R/I/A1/A2/E/B) for each methodological unit, as detailed in SKILL.md (lines 18-31).

3. Testing and Evolution Loop

Cangjie-skill automatically generates test-prompts.json during its pressure-test stage (lines 31-35), including bait tests and cross-skill confusion scenarios. Darwin-skill ingests these test prompts, executes them against the skill, and triggers revision cycles when failure rates exceed acceptable limits.

4. Ecosystem Integration

The pipeline flows sequentially: nuwa-skill → cangjie-skill → darwin-skill. Cangjie-skill's output is directly consumable by darwin-skill, enabling continuous improvement of book-derived skills without manual intervention.

The RIA-TV++ Pipeline Implementation

Cangjie-skill's differentiation from nuwa-skill lies in its rigorous RIA-TV++ methodology, documented in methodology/00-overview.md. The pipeline executes seven distinct stages:

1. Adler Analysis – Structural decomposition of the source text
2. Parallel Extraction – Simultaneous mining of methods and frameworks
3. Triple Verification – Cross-validation of extracted concepts
4. RIA++ Construction – Building the six-section skill units (R/I/A1/A2/E/B)
5. Zettelkasten Linking – Creating knowledge graph connections between atomic skills
6. Pressure Testing – Generating adversarial test cases including bait questions
7. Delivery – Packaging into the final skill directory structure

This structured approach contrasts with nuwa-skill's focus on stylistic mimicry and darwin-skill's optimization loops.

Practical Implementation Examples

The following commands illustrate the distinct roles each tool plays in the workflow:

# Distill a persona using nuwa-skill

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

# Convert a book to skills using cangjie-skill

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

# Evolve the skill using darwin-skill

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

Key Repository Files and Specifications

File Path	Purpose in the Ecosystem
SKILL.md	Defines the meta-skill specification and documents the relationship between the three tools, referencing nuwa-skill and darwin-skill in the ecosystem positioning section (lines 54-61)
README.en.md	Provides the high-level architectural narrative and ecosystem diagram linking the three skills (lines 7-9)
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

• Nuwa-skill converts human expression into mimicable agent personas by analyzing personal text outputs.
• Cangjie-skill transforms books and long-form content into structured, testable skill libraries using the RIA-TV++ seven-stage pipeline.
• Darwin-skill provides the evolution layer, automatically improving any skill through iterative testing against generated prompt suites.
• The three tools form a continuous pipeline: distill (nuwa/cangjie) → structure (cangjie) → test (cangjie) → evolve (darwin).
• Cangjie-skill's test-prompts.json output serves as the critical interface between the distillation and evolution layers.

Frequently Asked Questions

Can darwin-skill improve skills created by nuwa-skill?

Yes. Darwin-skill is designed to evolve any existing skill, including those produced by nuwa-skill. It consumes the SKILL.md and associated test prompts from nuwa-skill's output and applies the same automated feedback loops used for book-derived skills, continuously refining the persona's execution accuracy.

What distinguishes cangjie-skill's output structure from nuwa-skill?

While both produce SKILL.md files compatible with Claude, cangjie-skill follows the RIA-TV++ framework with six specific sections (R/I/A1/A2/E/B) for each atomic skill, as defined in templates/SKILL.md.template. Nuwa-skill focuses on stylistic and heuristic patterns rather than methodological frameworks, resulting in a different internal schema optimized for personality mimicry rather than procedure execution.

How does the pressure-testing stage work in cangjie-skill?

During the pressure-test stage, cangjie-skill generates test-prompts.json containing bait questions designed to trigger hallucinations and cross-skill confusion tests to verify boundary clarity. This file serves as the quality assurance input for darwin-skill's evolution cycles, ensuring that skills only activate within their intended operational domains.

Do these three skills require each other to function?

No. Each skill operates independently: nuwa-skill can create personas without cangjie-skill, and cangjie-skill can generate book skills without darwin-skill. However, when used together as described in README.en.md, they form a complete distill → structure → test → evolve pipeline that maximizes skill reliability and performance.

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