How to Build and Review OpenClaw Skills
Learn how to create custom OpenClaw skills, review community skills for quality and security, and decompose complex tasks into reusable skill chains.
Última atualização: 2026-03-31
What You'll Build
A skill development and review workflow that lets you:
- Build custom skills — create new OpenClaw skills tailored to your specific needs
- Review community skills — vet skills from ClawHub for quality, security, and reliability
- Decompose complex tasks — break multi-step workflows into reusable skill chains
- Iterate and improve — test, debug, and refine skills with AI-assisted feedback
By the end of this guide, you'll be able to create production-quality skills and confidently evaluate any skill before installing it.
Why Build Custom Skills
ClawHub has hundreds of community skills, but sometimes you need something specific:
- Internal tools — your company's APIs, dashboards, and workflows aren't covered by public skills
- Custom integrations — you use niche tools or proprietary systems that need custom connectors
- Workflow optimization — combining multiple skills into a single, streamlined skill saves time
- Quality control — building your own skills means you know exactly what they do
- Competitive advantage — unique automation workflows that your competitors don't have
Why Review Skills Before Installing
Not every skill on ClawHub is production-ready. Before installing a community skill:
- Security — a skill could access your file system, network, or credentials in ways you don't expect
- Quality — some skills are prototypes or proof-of-concepts, not production-grade
- Maintenance — abandoned skills may not work with newer OpenClaw versions
- Dependencies — a skill might require external services or API keys you don't have
- Permissions — understand exactly what access a skill requests before granting it
Prerequisites
- OpenClaw installed and configured (Getting Started Guide)
- Familiarity with YAML and basic scripting
- Node.js 18+
Step 1: Install the Skill Development Tools
# 1. Skill builder — guides you through creating well-structured skills npx clawhub@latest install skill-builder # 2. Skill reviewer — quality and structure analysis npx clawhub@latest install skill-reviewer # 3. Task decomposer — break complex tasks into skill chains npx clawhub@latest install task-decomposer
Step 2: Create Your First Custom Skill
Understand Skill Structure
An OpenClaw skill is centered around a SKILL.md file — this is the core instruction document that tells the AI agent how to perform a task. The Skill Builder skill guides you through creating well-structured skills following best practices.
A typical skill structure looks like:
my-custom-skill/ ├── SKILL.md # Core instructions for the agent ├── [topic].md # On-demand detail files (loaded when needed) ├── references/ # Heavy documentation, examples └── scripts/ # Optional automation scripts
Write Your SKILL.md
The SKILL.md is the heart of your skill. Ask OpenClaw (with the Skill Builder skill installed) to help you write one:
Help me create a skill that searches a custom API for products. The API is at https://api.example.com/search and requires an API key.
The Skill Builder guides you through:
- Modular structure — keeping the main SKILL.md focused and using separate files for detailed topics
- Progressive disclosure — loading detail files only when the agent needs them
- Token-efficient design — minimizing context window usage
- Clear input/output — defining what the skill expects and returns
Example SKILL.md
# Product Search Skill
Search the product catalog API for items matching a query.
## Configuration
- `API_URL`: Base URL of the product API
- `API_KEY`: API key for authentication (set as environment variable)
## Usage
When the user asks to search for products, use the API:
GET {API_URL}/search?q={query}
Authorization: Bearer {API_KEY}
## Response Format
Return results as a structured list with name, price, and URL for each product.
Test the Skill
Install your skill locally and test it:
clawhub install ./my-custom-skill
Then try using it in OpenClaw to verify it works as expected.
Step 3: Review a Community Skill
Before installing any skill from ClawHub, review it:
clawhub run skill-reviewer --skill "clawhub-username/skill-name"
The reviewer analyzes the skill's SKILL.md and supporting files, checking for:
=== Skill Review: awesome-slack-bot === ## Structure Assessment ✓ SKILL.md is well-organized with clear sections ✓ Description is specific and actionable ✓ Input/output expectations are documented ✓ Usage examples included ⚠ Missing on-demand detail files — all content is in one large SKILL.md ## Quality Score ✓ Frontmatter is complete (name, description, version) ✓ Instructions are token-efficient — progressive disclosure used ✓ Error handling guidance included ⚠ No rate limiting guidance — may hit API limits under heavy use ## Description Quality: 8.5/10 Recommendation: Well-structured skill, ready for use
What to Look For
Red Flags
- Skill requests permissions it shouldn't need (e.g., a search skill requesting file write access)
- No tests or documentation
- Last updated more than 6 months ago
- Hardcoded URLs or credentials in source code
- Obfuscated or minified code
Green Flags
- Clear permission declarations matching actual usage
- Comprehensive test coverage
- Active maintenance and community engagement
- Transparent source code
- Well-documented configuration
Step 4: Decompose Complex Tasks
The Task Decomposer skill helps break complex requests into executable steps. Ask OpenClaw:
Decompose this task: Monitor competitor pricing pages daily and alert me when prices change
Example output:
=== Task Decomposition === Task: Monitor competitor pricing pages daily and alert when prices change ## Skill Chain 1. **browser-use** — Navigate to each competitor's pricing page Input: List of competitor URLs Output: Page content / screenshots 2. **web-scraper** — Extract pricing data from each page Input: Page content Output: Structured pricing data (JSON) 3. **file_read** — Load yesterday's pricing data for comparison Input: File path to stored data Output: Previous pricing data 4. **summarize** — Compare current vs. previous pricing and identify changes Input: Current data + previous data Output: Change report 5. **telegram** (or email) — Send alert if changes detected Input: Change report Output: Notification sent ## Required Skills - browser-use (installed: yes) - web-scraper (installed: yes) - summarize (installed: yes) - telegram (installed: no — install with: npx clawhub@latest install telegram) ## Estimated Setup Time: 20 minutes
From Decomposition to Workflow
Turn the decomposition into a reusable workflow:
# .openclaw/price-monitor.yml
name: competitor-price-monitor
schedule: "0 9 * * *"
steps:
- skill: browser-use
action: navigate
urls:
- "https://competitor1.com/pricing"
- "https://competitor2.com/pricing"
- skill: web-scraper
action: extract
format: json
target: "pricing tables"
- skill: summarize
action: compare
current: "{{step_2.output}}"
previous: "./data/last-pricing.json"
- skill: telegram
action: send
condition: "changes_detected"
message: "{{step_3.output}}"
- action: save
data: "{{step_2.output}}"
path: "./data/last-pricing.json"
Step 5: Publish Your Skill
When your skill is ready, publish it to ClawHub:
clawhub publish my-custom-skill
Before publishing, ensure:
- SKILL.md is complete and well-structured
- Description is specific and actionable
- No hardcoded secrets in any files
- Version number follows semver
Troubleshooting
Skill not working as expected
- Check that SKILL.md instructions are clear and specific
- Test with different prompts to see if the agent interprets the skill correctly
- Review the Skill Builder guidelines for common instruction pitfalls
"Permission denied" during skill execution
- Review what the skill accesses — files, network, credentials
- Ensure you've granted the required permissions when installing
- Some permissions require explicit user approval on first run
Skill reviewer shows false positives
- The reviewer is conservative — it flags anything that could be a concern
- Review the specific warnings and decide if they apply to your use case
- Not every warning is a real problem — use your judgment
Published skill not appearing on ClawHub
- Publishing can take a few minutes to propagate
- Verify your ClawHub account is verified
- Check that the skill name doesn't conflict with an existing skill
Perguntas Frequentes
No. Most OpenClaw skills are built around a `SKILL.md` file — a Markdown document that instructs the AI agent how to perform a task. You can create effective skills with just clear written instructions and API documentation. For more complex automation, you can include scripts (Python, Bash, etc.) alongside the SKILL.md, but this is optional.
Yes. Skills can be installed locally without publishing. Keep the skill directory in your project or a private repository, and install it with a local path: `clawhub install ./my-custom-skill`. This is common for company-internal integrations that shouldn't be public.
Bump the version in your skill's metadata, update the changelog in README, and run `clawhub publish` again. ClawHub handles versioning — users on older versions won't be automatically updated unless they run `clawhub update`.
Follow the principle of least privilege. Only request permissions your skill actually needs. Common permission combinations: network-only (for API integrations), network + file_read (for data processing), network + credentials (for authenticated APIs). Avoid requesting file_write unless your skill needs to save output to disk.
Include rate limiting guidance in your SKILL.md — instruct the agent to implement exponential backoff when making API calls. For skills that interact with rate-limited APIs, document the limits in your README so users know what to expect, and include retry logic in any included scripts.
Yes — this is one of the most powerful patterns. Use YAML workflow definitions to chain skills together, or use the Task Decomposer to identify the right skill chain for your use case. The resulting workflow can be saved as a new skill that internally delegates to other skills. This is how most advanced automations are built.
Casos de Uso Relacionados
Automatizar Revisões de PR no GitHub
Configure um fluxo de revisão de PR automatizado com análise de código por IA, convenções de commits e feedback acionável.
Workflow de Pesquisa Pessoal
Construa um assistente de pesquisa com IA que busca na web, encontra artigos acadêmicos e sintetiza descobertas.
Automação de Navegador
Automatize tarefas do navegador com IA: web scraping, preenchimento de formulários, capturas de tela e fluxos web complexos usando linguagem natural.