How cangjie‑skill Works with darwin‑skill for Automatic Skill Evolution

cangjie‑skill emits a darwin‑compatible test‑prompts.json file that darwin‑skill consumes to automatically test, compare, and evolve the distilled book skill without manual retraining.

The cangjie‑skill repository implements a "book‑distilling" capability that converts written content into structured AI skills. To keep these skills current as underlying models improve, the project deliberately integrates with darwin‑skill—a generic evolution framework. This article explains the precise relationship between cangjie‑skill and darwin‑skill for skill evolution, based on the source code in kangarooking/cangjie-skill.

The Compositional Architecture: Distill, Then Evolve

The relationship between cangjie‑skill and darwin‑skill is compositional rather than dependent. Each tool has a single responsibility:

• cangjie‑skill performs the one-time work of ingesting a book and producing a skill directory.
• darwin‑skill handles the ongoing work of testing that skill against live model outputs and rewriting it when drift or improvement is detected.

This separation mirrors the broader ecosystem pattern described in README.en.md【Line 13‑Line 158】, where nuwa‑skill distills people-centric skills and darwin‑skill evolves them. Cangjie‑skill extends this pattern to book-centric skills.

Stage 5: Emitting Darwin‑Compatible Test Prompts

The five-stage pipeline in SKILL.md【Line 28‑Line 64】culminates in a delivery stage that produces evolution-ready artifacts. Specifically, cangjie‑skill generates test‑prompts.json with a mandatory darwin_compatible: true flag.

The template at templates/test‑prompts.json.template【Line 5】defines this format:

{
  "darwin_compatible": true,
  "test_cases": [
    {
      "prompt": "Summarize Chapter 3's main argument",
      "expected_behavior": "Identifies the thesis about economic incentives"
    }
  ]
}

This structured format allows darwin‑skill to parse the file unambiguously and execute the defined prompts against the current model version.

The Evolution Command Chain

According to methodology/07‑stage5‑deliver.md【Line 61】, the handoff from cangjie‑skill to darwin‑skill uses a simple directory path convention:

# 1️⃣ Run cangjie‑skill on a book (produces `test‑prompts.json`)

cangjie run --book my‑great‑novel --output-dir books/my‑great‑novel

# 2️⃣ Feed the resulting skill to darwin‑skill for automatic evolution

darwin evolve books/my‑great‑novel/

The darwin evolve command performs three operations:

1. Reads test‑prompts.json from the provided path.
2. Executes each prompt against the current model.
3. Compares outputs to the stored baseline; if improvements exceed the threshold, rewrites the skill files.

This enables continuous evolution without requiring the original book or manual prompt engineering.

Key Files Supporting the Integration

File	Purpose
templates/test‑prompts.json.template	Defines the darwin‑compatible schema with darwin_compatible: true【templates/test‑prompts.json.template†Line 5】
SKILL.md	Documents the five-stage pipeline including the pressure‑test stage【SKILL.md†Line 28‑Line 64】
README.en.md	Explains ecosystem relationships between nuwa‑skill, cangjie‑skill, and darwin‑skill【README.en.md†Line 13‑Line 158】
methodology/07‑stage5‑deliver.md	Specifies the darwin evolve command for skill evolution【methodology/07‑stage5‑deliver.md†Line 61】

WhyThis Design Matters

The cangjie‑skill and darwin‑skill integration solves a critical maintenance problem: model drift. As base language models improve, a skill distilled six months ago may produce suboptimal outputs even if its logic remains sound. By externalizing test cases into a darwin‑compatible format, cangjie‑skill ensures that:

• Evolution happens automatically in CI pipelines.
• Test coverage remains portable across model versions.
• Skills stay current without reprocessing the source book.

Summary

• cangjie‑skill and darwin‑skill operate as compositional tools in a distill‑then‑evolve pipeline.
• cangjie‑skill emits test‑prompts.json with darwin_compatible: true【templates/test‑prompts.json.template†Line 5】.
• darwin‑skill consumes this file via darwin evolve <path>【methodology/07‑stage5‑deliver.md†Line 61】 to test and rewrite skills automatically.
• This architecture enables continuous skill maintenance without manual intervention or source material reprocessing.

Frequently Asked Questions

Does cangjie‑skill require darwin‑skill to function?

No. cangjie‑skill produces a complete, runnable skill directory independently. darwin‑skill is optional and used only when you want automated evolution. The test‑prompts.json file is safely ignored if darwin‑skill is not present.

What happens if darwin‑skill finds no improvements during evolution?

darwin‑skill leaves the skill files unchanged and reports the comparison results. No rewrite occurs unless the new model outputs exceed the configured improvement threshold, protecting against regression.

Can I use darwin‑skill with skills that weren't created by cangie‑skill?

Yes. darwin‑skill is a generic framework that can evolve any skill providing a valid test‑prompts.json. The cangjie‑skill integration simply automates the creation of that file in the correct format【README.en.md†Line 13‑Line 158】.

How does this relate to nuwa‑skill in the broader ecosystem?

nuwa‑skill distills people (interviews, profiles) into skills; cangjie‑skill distills books. Both emit darwin‑compatible artifacts so darwin‑skill can evolve either type of skill automatically【README.en.md†Line 13‑Line 158】.

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