How cangjie-skill, nuwa-skill, and darwin-skill Form an Integrated Ecosystem for Knowledge Distillation

cangjie-skill, nuwa-skill, and darwin-skill work together as a closed-loop pipeline that distills expertise from both people and books, then continuously evolves those skills through automated testing and improvement.

The cangjie-skill integration ecosystem enables a complete "distillation-evolution" workflow for transforming human knowledge into maintainable, testable AI skills. This article examines how three complementary repositories interlock to create a self-sustaining knowledge pipeline, with specific details from the kangarooking/cangjie-skill source code.

The Three Pillars of the Ecosystem

Each project serves a distinct purpose in the knowledge lifecycle:

Component	Responsibility	Output
nuwa-skill	Distills people — thinking patterns, expression DNA, personal expertise	Human skill artifact (e.g., "Elon Musk skill")
cangjie-skill	Distills books — methodologies, frameworks, principles, actionable guidance	Book-derived skill with test-prompts.json
darwin-skill	Evolves any skill through continuous testing and feedback loops	Iteratively improved skill definitions

As documented in SKILL.md lines 54-60:

"三者咬合: 本 skill 输出的 test-prompts.json 严格遵循 darwin-skill 格式, 以便产出的 skill 可直接接入 darwin 做自动进化。"

This three-way interlock ensures that outputs from both distillation engines feed seamlessly into the evolution engine.

How the Integration Ecosystem Works: The Four-Phase Pipeline

Phase 1: Extraction

Either nuwa-skill (for people) or cangjie-skill (for books) parses source material and produces structured artifacts:

• skill.json — core skill definition
• test-prompts.json — evaluation test cases
• DIGEST.md — human-readable summary (from templates/DIGEST.md.template)

Phase 2: Validation

cangjie-skill enforces "Triple Verification" before export. The methodology in methodology/07-stage5-deliver.md specifies quality gates including:

• R/I/A1/A2/E/B section completeness
• Citation limit compliance
• Schema conformance to darwin-skill expectations

Phase 3: Packaging

Finalized skills land in standard locations:

• User level: ~/.claude/skills/
• Project level: .claude/skills/

The delivery stage documented in methodology/07-stage5-deliver.md handles this placement and signals readiness for evolution.

Phase 4: Evolution

darwin-skill monitors the skill repository, runs automated test suites, and iteratively refines prompts and test coverage. This closes the loop: new books or people models continuously feed into an ever-improving skill library.

Practical Integration: Command-Line Workflow

The following demonstrates the complete cangjie-skill integration ecosystem in operation:

# 1️⃣ Distill a book into a skill (cangjie-skill)

cangjie-skill distill \
  --source-url "https://example.com/my-book.pdf" \
  --output-dir "$HOME/.claude/skills/my-book-skill"

# 2️⃣ (Optional) Distill a person into a skill (nuwa-skill)

nuwa-skill distill \
  --profile "elon-musk" \
  --output-dir "$HOME/.claude/skills/elon-musk-skill"

# 3️⃣ Run darwin-skill to evolve the newly created skill

darwin-skill evolve \
  --skill-dir "$HOME/.claude/skills/my-book-skill"

Key integration points:

• cangjie-skill distill outputs test-prompts.json in darwin-skill's expected schema
• nuwa-skill distill produces compatible artifacts for person-derived expertise
• darwin-skill evolve consumes either source type indiscriminately

Schema Compliance: The Critical Link

The cangjie-skill integration ecosystem depends on strict output formatting. According to SKILL.md, cangjie-skill's generated test-prompts.json must strictly follow darwin-skill's format specification. This contract enables:

• Automatic ingestion without transformation
• Consistent evaluation across skill origins
• Portable evolution regardless of source (person vs. book)

Key Source Files and Their Roles

File	Purpose	Location
SKILL.md	Ecosystem positioning, quality gates, three-skill relationship	Repository root
README.md	High-level ecosystem overview and usage entry points	Repository root
methodology/07-stage5-deliver.md	Final delivery orchestration, directory placement, darwin-skill handoff	methodology/
templates/DIGEST.md.template	Human-readable digest generation template	templates/
test-prompts.json (runtime)	Darwin-compatible test suite (generated per skill)	Skill output directory

Summary

• nuwa-skill distills people, cangjie-skill distills books, and darwin-skill evolves both
• The integration ecosystem relies on schema-compliant test-prompts.json as the universal interface
• Four phases (extraction → validation → packaging → evolution) create a closed improvement loop
• Standard directory conventions (~/.claude/skills/) enable seamless handoffs between components
• Source documentation in SKILL.md lines 54-60 and README.md lines 24-27 establish the architectural contract

Frequently Asked Questions

What makes cangjie-skill different from nuwa-skill?

cangjie-skill processes long-form written content (books, papers, documentation) to extract methodologies and frameworks, while nuwa-skill models individual people—their thinking patterns, communication style, and personal expertise. Both output skill artifacts, but their source materials differ fundamentally.

Can darwin-skill work with skills from only one source?

Yes. darwin-skill is source-agnostic. It evolves any skill that conforms to its expected schema, whether originating from cangjie-skill, nuwa-skill, or manual creation. The evolution engine operates on the standardized test-prompts.json format, not the provenance of the skill.

Where is the integration contract documented?

The three-skill relationship and schema requirements appear in SKILL.md (lines 54-60) within the kangarooking/cangjie-skill repository, with additional context in README.md (lines 24-27). The delivery-stage specifics are detailed in methodology/07-stage5-deliver.md.

What happens if test-prompts.json doesn't match darwin-skill's format?

The skill will fail to ingest into the evolution pipeline. cangjie-skill's validation phase enforces format compliance before export, preventing this failure mode. This quality gate ensures only compatible skills reach darwin-skill.

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