Skill Creator is a specialized meta-skill designed to help developers build, test, and refine new AI agent capabilities through an iterative, data-driven workflow.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install skill-creator-cn
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install skill-creator-cn using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Skill Creator?
Skill Creator serves as a comprehensive development environment for building Openclaw Skills. It guides users through a structured lifecycle of capturing intent, drafting technical instructions in Markdown, and performing rigorous evaluations. By leveraging subagents, it can simultaneously run test prompts with and without the new skill to provide a clear performance baseline.
The skill emphasizes high-quality instruction writing, advocating for a principle of lack of surprise and the use of theory of mind to help LLMs understand the why behind tasks. It also includes sophisticated tools for description optimization, ensuring that skills trigger accurately when faced with complex, real-world user queries.
Skill Creator Use Cases
- Creating a new AI agent skill from scratch based on a specific workflow or conversation history.
- Improving an existing skill by running comparative benchmarks against a snapshot of the old version.
- Optimizing skill descriptions to improve triggering accuracy and reduce under-triggering in Claude.
- Bundling repetitive manual steps into deterministic Python scripts stored within a skill's directory.
- Conducting blind A/B comparisons to objectively determine which skill iteration produces better qualitative results.
How Skill Creator Works
- Intent Capture: The agent interviews the user to define the skill's purpose, trigger contexts, and expected output formats.
- Drafting: A SKILL.md file is created with YAML frontmatter for metadata and Markdown for core instructions.
- Test Case Creation: The agent generates 2-3 realistic test prompts and saves them to a standardized JSON schema.
- Parallel Execution: Subagents run the test cases with the skill and a baseline (no skill or old version) to collect performance data.
- Evaluation: Quantitative assertions are graded, and a web-based Eval Viewer is launched for qualitative human review.
- Iteration: The skill is refined based on user feedback and benchmark data, repeating the loop until performance is stabilized.
- Optimization & Packaging: The skill description is optimized through a training loop, and the final result is packaged into a .skill file.
Skill Creator Setup
To use Skill Creator within your environment, ensure you have the necessary scripts available and run the following to package your progress:
# Package the skill for distribution
python -m scripts.package_skill <path/to/skill-folder>
# Run the description optimization loop
python -m scripts.run_loop --eval-set <path-to-eval.json> --skill-path <path-to-skill>
Skill Creator Data Schema & Taxonomy
Skill Creator uses a hierarchical directory structure and specific JSON schemas to manage the development lifecycle:
| File/Folder | Description |
|---|---|
SKILL.md |
The core definition file containing name, description, and instructions. |
evals/evals.json |
Stores test prompts, expected outputs, and quantitative assertions. |
scripts/ |
Directory for bundled executable scripts used by the skill. |
timing.json |
Captured for each run to track total_tokens and duration_ms. |
benchmark.json |
Aggregated data showing pass rates and deltas between iterations. |
Skill Creator Advanced Features
- Multi-Agent Benchmarking: Spawns parallel subagents to compare skill performance against baselines in real-time.
- Description Optimization Loop: Automatically iterates on frontmatter descriptions using a train/test split to maximize triggering accuracy.
- Eval Viewer: Generates a standalone HTML interface for side-by-side output comparison and feedback collection.
- Automated Grading: Uses specialized grader agents or Python scripts to verify assertions against model outputs.
- Blind Comparison: Supports rigorous A/B testing where an independent agent judges outputs without knowing which version produced them.
SKILL.md
Loading
Related Openclaw Skills
Workspace Local Retrieval
A professional framework for building local-first retrieval architectures with strict security boundaries and automated maintenance workflows.
Sendoso
A comprehensive integration for managing personalized gifting, eGifts, and direct mail campaigns through AI agents.
ChartGen AI
ChartGen AI is a comprehensive visualization and data analysis skill that transforms raw data into professional charts, dashboards, and PowerPoint reports.
Veo 3.1 Video Generation
A professional-grade video generation skill for creating cinematic, multi-shot AI videos via the NanoPhoto.AI Veo 3.1 API.
Mambo TTS
A high-performance text-to-speech skill for Openclaw providing lively, pre-configured Chinese voice profiles based on Edge TTS.
Spydra
Spydra is a high-performance web scraping and data extraction tool designed for researchers and businesses to programmatically collect web information.
Featured*
