AI Agents Compatible with Cangjie‑Skill Outputs: Claude Code, Cursor & OpenClaw

Cangjie‑Skill produces Claude‑compatible skill packs that work with Claude Code, Cursor, and OpenClaw, or any AI agent that can load a directory‑based skill structure.

Cangjie‑Skill is an open‑source framework that distills books into executable AI skills. Each skill is a self‑contained directory with a SKILL.md file and supporting assets. Because the format follows the Claude Code skill specification, multiple AI agents can consume these outputs natively.

What the "Cangjie‑Skill Output Format" Actually Is

Every generated skill follows a predictable structure:

• SKILL.md — the core definition with Research, Ideation, A1/A2 analysis, Execution, and Benchmark sections
• test-prompts.json — validation prompts to verify skill behavior
• Supporting files (templates, digests) for human readability

This directory‑of‑markdown design means no proprietary binary formats or vendor‑locked APIs. Any agent that implements the Claude skill directory protocol can load and execute these skills.

Verified AI Agents Compatible with Cangjie‑Skill

According to the [README.md](https://github.com/kangarooking/cangjie-skill/blob/main/README.md) and delivery documentation in [methodology/07-stage5-deliver.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/07-stage5-deliver.md), these three agents are explicitly supported:

Claude Code

The official Claude development environment and reference implementation for the skill format. Skills install to ~/.claude/skills/ and integrate directly into Claude's chat interface.

Key capability: Automatic trigger detection — Claude recognises when a user prompt matches an installed skill's scope and executes the E (Execution) steps automatically.

Cursor

The AI‑native IDE that reads Claude‑style skill folders from project‑local .cursor/skills/ directories. Allows per‑project skill customization without polluting global user settings.

Key capability: Skills coexist with codebase context, making them ideal for domain‑specific development workflows.

OpenClaw

A community‑maintained runtime that executes Claude skills outside official Anthropic infrastructure. Useful for self‑hosted or offline scenarios where Claude Code cannot run.

Key capability: CLI‑first execution with minimal dependencies, enabling CI/CD pipelines and automated testing of skills.

Installing a Cangjie‑Skill for Each Agent

The following commands reproduce the installation steps from [07-stage5-deliver.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/07-stage5-deliver.md) lines 45‑52.

Claude Code: User‑Level Installation

# Path to your generated skill directory

SKILL_DIR="books/my-book/awesome-skill"

# Target location in Claude's user skills path

TARGET="${HOME}/.claude/skills/awesome-skill"

# Copy or symlink the entire directory

cp -r "$SKILL_DIR" "$TARGET"

# Verify the skill loads correctly

claude skill test --skill "$TARGET" --prompt "Should trigger awesome-skill?"

Cursor: Project‑Level Installation

# Skill source directory

SKILL_DIR="books/my-book/awesome-skill"

# Project-local target (create if missing)

TARGET=".cursor/skills/awesome-skill"
mkdir -p "$TARGET"

# Copy skill contents

cp -r "$SKILL_DIR/"* "$TARGET/"

Once installed, invoke in any Cursor chat:

> ask: "How do I apply the principle from awesome-skill?"

Testing Skill Execution in Claude

After installation, Claude automatically detects relevant skills during conversation:

User: "Help me decide the best pricing model for my SaaS business."

Assistant (Claude): *[detects "pricing-model-skill" is installed]*
→ Executes the skill's Execution steps and returns a concrete plan.

Key Files That Define the Compatibility

File	Purpose	Relevance to Agents
[SKILL.md](https://github.com/kangarooking/cangjie-skill/blob/main/SKILL.md)	Core skill definition (R/I/A1/A2/E/B structure)	Required — parsed by all compatible agents
[methodology/07-stage5-deliver.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/07-stage5-deliver.md)	Delivery & installation instructions	Reference for host‑agent setup
templates/SKILL.md.template	Generator template for new skills	Defines output schema
[README.md](https://github.com/kangarooking/cangjie-skill/blob/main/README.md)	Platform badges and workflow overview	Documents verified agent compatibility

Custom Agents and Future Compatibility

Any AI agent that implements the following can consume Cangjie‑Skill outputs:

1. Directory traversal — load a folder of markdown files
2. SKILL.md parsing — extract the R/I/A1/A2/E/B sections
3. Trigger matching — decide when to activate based on user input
4. Execution routing — run the E section's defined steps

This includes custom agents built on Anthropic's skill API, LangChain tool definitions, or internal enterprise platforms. The format's simplicity deliberately avoids vendor lock‑in.

Summary

• Cangjie‑Skill outputs are Claude‑compatible skill directories, not proprietary packages
• Three agents are explicitly verified: Claude Code (user‑level), Cursor (project‑level), and OpenClaw (runtime)
• Installation is filesystem‑only — copy or symlink into ~/.claude/skills/ or .cursor/skills/
• Future agents can adopt the format by implementing the SKILL.md parsing protocol
• All compatibility documentation lives in [README.md](https://github.com/kangarooking/cangjie-skill/blob/main/README.md) and [methodology/07-stage5-deliver.md](https://github.com/kangarooking/cangjie-skill/blob/main/methodology/07-stage5-deliver.md)

Frequently Asked Questions

Does Cangjie‑Skill work with ChatGPT or GPT‑4?

No direct support. ChatGPT and GPT‑4 do not implement the Claude skill directory protocol. However, the SKILL.md content can be manually pasted into prompts or converted to custom GPT instructions. For native skill execution, use Claude Code, Cursor, or OpenClaw.

Can I use one skill across multiple agents simultaneously?

Yes. Copy the same skill directory to both ~/.claude/skills/ and .cursor/skills/ within a project. Both agents will load identical definitions. Consider using symlinks to keep versions synchronized during development.

What happens if an agent doesn't recognize a skill's trigger?

Claude Code and Cursor fall back to standard conversational behavior — the skill remains available for manual invocation by name. OpenClaw provides explicit CLI commands to force execution regardless of trigger matching. Test coverage via test-prompts.json helps validate trigger sensitivity.

Is OpenClaw officially maintained by Anthropic?

No. OpenClaw is community‑maintained and listed in the Cangjie‑Skill [README.md](https://github.com/kangarooking/cangjie-skill/blob/main/README.md) as a compatible third‑party runtime. It implements the same protocol but operates independently of Anthropic's official Claude Code distribution.

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