Cangjie-Skill vs Nuwa-Skill: Understanding the Distillation Differences in the kangarooking Ecosystem

cangjie-skill distills long-form content into atomic, test-validated skills while nuwa-skill distills human personas into single voice-mimicking skills.

Both tools belong to the kangarooking/cangjie-skill repository ecosystem but target fundamentally different source materials. cangjie-skill processes books, transcripts, and structured knowledge into a network of executable skills. nuwa-skill captures a person's thinking style and expression DNA into a unified skill package. Their architectural differences reflect these distinct goals.

What Each Skill Distills

cangjie-skill: Content-Centric Distillation

cangjie-skill focuses on "蒸馏书" (distilling books). It extracts methodologies, frameworks, principles, and actionable procedures from long-form content.

According to the repository's README.md, this tool handles:

• Books and academic papers
• Video transcripts and podcasts
• Online courses and structured documentation
• Any text-heavy material containing reusable methods

The output is a directory of atomic skills, each validated through rigorous verification stages.

nuwa-skill: Persona-Centric Distillation

nuwa-skill focuses on "蒸馏人" (distilling people). It captures human-centric attributes:

• Thinking style and reasoning patterns
• Expression DNA and phrasing habits
• Personal voice and communication patterns

The result is a single "human skill"—such as an "Elon Musk skill" or "Warren Buffett skill"—that reproduces how a specific individual talks and reasons.

Core Pipeline Architectures

The cangjie-skill RIA-TV++ Pipeline

In SKILL.md, the cangjie-skill pipeline spans seven structured stages:

1. Adler-style whole-book understanding — comprehensive structural analysis
2. Parallel extraction — simultaneous framework, principle, case, counter-example, and glossary extraction
3. Triple Verification — cross-domain evidence, predictive power, and uniqueness validation
4. RIA++ skill construction — building Reading, Interpretation, Application, and Execution components
5. Zettelkasten linking — creating networked knowledge relationships
6. Pressure-testing — darwin-compatible validation
7. Final delivery — packaged skills ready for consumption

This pipeline definition appears at lines 22-30 of SKILL.md in the repository.

The nuwa-skill Simplified Pipeline

nuwa-skill employs a content-agnostic text-generation approach as described in README.md (lines 24-27). The pipeline emphasizes:

• Reproducing persona voice without method-level decomposition
• Wrapping output as Claude-compatible skills
• Delegating evolution to darwin-skill post-creation

The reduced complexity reflects the focused goal: mimicry rather than methodological extraction.

Output Granularity and Structure

cangjie-skill: Atomic Skill Networks

Each cangjie-skill distillation produces:

• SKILL.md files with standardized R/I/A1/A2/E/B sections:
  • R — Reading (source material reference)
  • I — Interpretation (core concept explanation)
  • A1 — Past Application (historical use cases)
  • A2 — Future Trigger (activation conditions)
  • E — Execution (actionable steps)
  • B — Boundary (limitations and edge cases)

These sections are defined at lines 54-60 of SKILL.md. The template lives in templates/SKILL.md.template.

Additionally, cangjie-skill generates test-prompts.json files for automatic evolution through darwin-skill.

nuwa-skill: Unified Persona Packages

nuwa-skill outputs contain:

• A single SKILL.md encapsulating overall expression patterns
• Example prompts demonstrating the captured voice
• No decomposition into multiple method-level skills

The structure prioritizes holistic persona preservation over granular reusability.

Verification and Quality Assurance

Triple Verification in cangjie-skill

The cangjie-skill pipeline enforces Triple Verification (lines 98-105 of SKILL.md):

Verification Type	Purpose
Cross-domain evidence	Validates applicability beyond source context
Predictive power	Tests whether the skill generates useful forecasts
Uniqueness	Ensures non-redundancy with existing skills

Pressure testing ensures darwin-skill compatibility before delivery.

nuwa-skill: Data-Dependent Quality

nuwa-skill relies on:

• Quality and volume of original persona data
• Post-hoc evolution through darwin-skill
• No built-in verification equivalent to Triple Verification

This trade-off accepts higher variance in exchange for simpler operation.

Ecological Positioning

The README.md (lines 54-58) positions these tools within a three-part ecosystem:

Tool	Role	Metaphor
cangjie-skill	Distill structured knowledge	"蒸馏书" (distill books)
nuwa-skill	Distill human personas	"蒸馏人" (distill people)
darwin-skill	Evolve any skill	Evolution engine

Both distillation outputs feed into darwin-skill for refinement, but their entry points differ dramatically.

Practical Invocation Examples

Using cangjie-skill

请帮我把《穷查理宝典》蒸馏成 skill。

Result structure:

books/poor-charlies-almanack/
├── BOOK_OVERVIEW.md          # Adler-style structural analysis

├── SKILL.md                  # Multiple atomic skills with R/I/A1/A2/E/B

├── INDEX.md                  # Zettelkasten-style cross-references

└── test-prompts.json         # Darwin-compatible pressure tests

Using nuwa-skill

把 Elon Musk 的表达方式蒸馏成 skill。

Result structure:

elon-musk-skill/
└── SKILL.md                  # Unified persona capture with example prompts

Key Implementation Files

Understanding the distillation difference requires familiarity with these repository files:

• README.md — Ecosystem overview and tool relationships
• SKILL.md — Master meta-skill with RIA-TV++ pipeline definition
• methodology/00-overview.md — Stage-by-stage pipeline documentation
• extractors/framework-extractor.md — Parallel extraction prompt templates
• templates/SKILL.md.template — Atomic skill generation skeleton
• templates/test-prompts.json.template — Darwin-skill test blueprint

Summary

• cangjie-skill processes long-form content through a seven-stage RIA-TV++ pipeline, producing multiple atomic, verified skills with structured R/I/A1/A2/E/B sections and built-in pressure testing.
• nuwa-skill applies a simpler, content-agnostic pipeline to human sources, generating single unified skills that capture expression DNA and reasoning patterns.
• Both tools complement each other in the kangarooking ecosystem—cangjie-skill for methodological knowledge, nuwa-skill for persona mimicry—with darwin-skill handling subsequent evolution.

Frequently Asked Questions

Can cangjie-skill and nuwa-skill be used on the same source material?

Generally no—cangjie-skill expects structured, method-bearing content while nuwa-skill requires human-centric expression samples. A book about a person might feed cangjie-skill; that person's interviews and writings would feed nuwa-skill. The tools are architecturally separated because their extraction targets differ fundamentally.

Why does cangjie-skill produce multiple files while nuwa-skill produces one?

Source material determines output structure. Books contain dozens of distinct methodologies worth isolating, hence atomic skills. A person's expression style is inherently unified—splitting it would fragment the very voice being captured. The repository's SKILL.md (lines 54-60) explicitly defines this granularity difference.

Are the skills from both tools compatible with Claude?

Yes. Both output Claude-compatible skill packages, though cangjie-skill includes additional darwin-skill test hooks. The README.md notes that nuwa-skill wraps output "as a Claude-compatible skill" while cangjie-skill template files ensure the same compatibility with extended validation layers.

How does darwin-skill interact with these distillation tools?

darwin-skill serves as the evolution layer. It accepts outputs from either cangjie-skill or nuwa-skill and refines them through automated testing. cangjie-skill includes pre-packaged test prompts (test-prompts.json) for this purpose; nuwa-skill delegates testing entirely to the darwin stage.

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