Cangjie-Skill vs Nuwa-Skill vs Darwin-Skill: The Complete Skill Ecosystem Guide

Cangjie-skill, nuwa-skill, and darwin-skill form a three-layer skill ecosystem that together enable AI agents to distill, structure, and continuously evolve knowledge from different sources.

This guide breaks down the relationship between these three open-source tools developed by kangarooking. Each tackles a distinct knowledge source—people, books, and existing skills—and produces executable AI capabilities through a unified pipeline.

What Each Skill Tool Does

The three tools occupy different layers in a knowledge processing stack. Understanding their individual roles clarifies how they interlock.

Nuwa-Skill: Distilling Human Personas

Nuwa-skill transforms person-centric content into callable AI skills. It extracts what makes someone think distinctively—their expression patterns, decision heuristics, and experiential worldview.

The tool takes inputs like tweets, interview transcripts, or speeches and produces a Claude-compatible SKILL.md. Instead of generic responses, the resulting skill mimics the target person's reasoning style and communication DNA.

According to the cangjie-skill README, nuwa-skill handles the "people" layer of the ecosystem【/cache/repos/github.com/kangarooking/cangjie-skill/main/README.en.md†L7-L9】.

Cangjie-Skill: Distilling Books and Long-Form Content

Cangjie-skill converts books, courses, podcasts, and lengthy documents into atomic, executable skills. Unlike nuwa-skill’s persona focus, cangjie-skill extracts methodological frameworks and makes them operable.

The tool implements a RIA-TV++ pipeline visible in SKILL.md:

1. Adler analysis – structural decomposition of the source material
2. Parallel extraction – multi-angle content harvesting
3. Triple verification – cross-checking for accuracy
4. RIA++ construction – building the six-section skill format
5. Zettelkasten linking – connecting related concepts
6. Pressure testing – validating with test prompts
7. Delivery – packaging the final skill set

Each atomic skill follows a rigid SKILL.md structure with six sections: Reference, Instruction, Application 1, Application 2, Evaluation, and Boundary conditions【/cache/repos/github.com/kangarooking/cangjie-skill/main/SKILL.md†L18-L31】.

The output includes:

• BOOK_OVERVIEW.md – holistic source summary
• INDEX.md – navigable skill directory
• GLOSSARY.md – defined terms and concepts
• Individual SKILL.md files – one per atomic method
• test-prompts.json – validation cases for darwin-skill

Darwin-Skill: Continuous Skill Evolution

Darwin-skill provides automatic improvement for any existing skill—including those produced by nuwa-skill and cangjie-skill. It closes the loop by feeding real performance data back into skill refinement.

The tool consumes test-prompts.json (generated during cangjie-skill's pressure-test stage) and runs an iterative feedback cycle:

• Execute test prompts against the skill
• Evaluate output quality
• Rewrite prompts, boundaries, and execution steps when performance degrades
• Repeat until convergence or iteration limit

The README describes darwin-skill as the "companion" that evolves any skill in the ecosystem【/cache/repos/github.com/kangarooking/cangjie-skill/main/README.en.md†L24-L30】.

How the Three Tools Connect

The relationship between cangjie-skill, nuwa-skill, and darwin-skill follows a clear data flow:

Stage	Tool	Input	Output
1. Distillation	nuwa-skill OR cangjie-skill	Person text (tweets, interviews) OR Book text (PDFs, transcripts)	SKILL.md + metadata
2. Testing	cangjie-skill (internal)	Extracted methods	test-prompts.json
3. Evolution	darwin-skill	Any skill + test prompts	Improved skill definition

Key Architectural Differences

Nuwa-skill emphasizes style and tone. Its output captures how a person would say something, making it ideal for creative tasks, advice simulation, or communication assistance.

Cangjie-skill emphasizes actionable methodology. Its RIA-TV++ pipeline forces every extracted concept into an executable format with clear boundaries and evaluation criteria. This suits analytical tasks, decision frameworks, and procedural knowledge.

Darwin-skill is source-agnostic. Whether the input came from a person or a book, it applies the same evolutionary pressure to improve robustness over time.

Practical Usage Example

Here's how the three tools combine in a real workflow:

# Step 1: Capture Charlie Munger's thinking style

nuwa-skill --source ./munger_interviews.txt \
           --profile "value_investing_persona" \
           --output ./skills/munger-persona

# Step 2: Distill his book into executable methods

cangjie-skill --text ./poor_charlies_almanack.txt \
              --title "Poor Charlie's Almanack" \
              --author "Charlie Munger" \
              --output ./skills/munger-methods

# Step 3: Continuously improve both skills

darwin-skill --skill-dir ./skills/munger-persona \
             --test-cases ./skills/munger-persona/test-prompts.json \
             --iterations 10

darwin-skill --skill-dir ./skills/munger-methods \
             --test-cases ./skills/munger-methods/test-prompts.json \
             --iterations 10

The commands above illustrate conceptual usage patterns based on the ecosystem architecture described in cangjie-skill's documentation.

Critical Files in Cangjie-Skill

These files define the integration points with nuwa-skill and darwin-skill:

• SKILL.md – Core specification including ecosystem positioning and RIA-TV++ pipeline details【/cache/repos/github.com/kangarooking/cangjie-skill/main/SKILL.md†L54-L60】
• README.en.md – Ecosystem overview with explicit links to nuwa-skill and darwin-skill【/cache/repos/github.com/kangarooking/cangjie-skill/main/README.en.md†L7-L9】
• methodology/00-overview.md – Seven-stage pipeline documentation
• templates/SKILL.md.template – R/I/A1/A2/E/B structure template

Summary

• Nuwa-skill distills people into persona-based skills via style extraction
• Cangjie-skill distills books into method-based skills via the RIA-TV++ pipeline
• Darwin-skill evolves any skill through automated testing and iterative refinement
• Together they form a distill → structure → test → evolve pipeline for AI-native knowledge tools
• Cangjie-skill auto-generates test-prompts.json specifically to feed darwin-skill's evolution loop

Frequently Asked Questions

Can I use darwin-skill with skills I created manually?

Yes. Darwin-skill is source-agnostic. As long as your skill follows Claude's skill schema and includes test cases (or you provide them separately), darwin-skill can evolve it. It was designed specifically to improve outputs from both nuwa-skill and cangjie-skill, but works with any compatible skill definition.

What makes cangjie-skill's RIA-TV++ pipeline different from standard summarization?

RIA-TV++ forces actionability. Standard summarization produces information you can read. Cangjie-skill produces skills you can invoke. The six-section structure (R/I/A1/A2/E/B) in every SKILL.md guarantees each extracted method has clear instructions, dual application examples, evaluation criteria, and explicit boundaries—none of which are guaranteed by conventional summarization.

Why separate nuwa-skill and cangjie-skill instead of one unified tool?

The knowledge sources require fundamentally different extraction strategies. Persona distillation prioritizes linguistic patterns and decision heuristics from unstructured biographical content. Book distillation prioritizes methodological decomposition from structured long-form arguments. Combining them would force compromises in both. The separation allows each tool to optimize for its specific input type while producing compatible outputs for darwin-skill's unified evolution layer.

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