Design and build OpenClaw skills. Use when asked to "make/build/craft a skill", extract ad-hoc functionality into a skill, or package scripts/instructions for reuse. Covers OpenClaw-specific integrati
Real data. Real impact.
Emerging
Developers
Per week
Open source
Skills give you superpowers. Install in 30 seconds.
An opinionated guide for creating OpenClaw skills. Focuses on OpenClaw-specific integration — message routing, cron scheduling, memory persistence, channel formatting, frontmatter gating — not generic programming advice.
Docs: https://docs.openclaw.ai/tools/skills · https://docs.openclaw.ai/tools/creating-skills
This skill is written for frontier-class models (Opus, Sonnet). If you're running a cheaper model and find a stage underspecified, expand it yourself — the design sequence is a scaffold, not a script. Cheaper models should:
{baseDir}/patterns/Skip if building from scratch. Use when packaging existing functionality (scripts, TOOLS.md sections, conversation patterns, repeated instructions) into a skill.
Gather what exists, where it lives, what works, what's fragile. Then proceed to Stage 1.
Work through with the user:
Ask early: Is this for your setup, or should it work on any OpenClaw instance?
| Choice | Implications |
|---|---|
| Universal | Generic paths, no local assumptions, ClawHub-ready |
| Particular | Can reference local skills, tools, workspace config |
Scan
<available_skills>Review the docs with the skill's needs in mind. Think compositionally — OpenClaw's primitives combine in powerful ways. Key docs to check:
| Need | Doc |
|---|---|
| Messages | |
| Cron/scheduling | |
| Subagents | |
| Browser | |
| Canvas UI | |
| Node devices | |
| Slash commands | |
See
{baseDir}/patterns/composable-examples.mdBased on Stages 1–2, identify which patterns apply:
| If the skill... | Pattern |
|---|---|
| Wraps a CLI tool | |
| Wraps a web API | |
| Monitors and notifies | |
Load all that apply and synthesise. Most skills combine patterns.
Script vs. instructions split: Scripts handle deterministic mechanics (API calls, data gathering, file processing). SKILL.md instructions handle judgment (interpreting results, choosing approaches, composing output). The boundary is: could a less intelligent system do this reliably? If yes → script.
Present proposed architecture for user review:
State locations:
<workspace>/memory/{baseDir}/state.json<workspace>/state/<skill>.jsonIf extracting: include migration notes (what moves, what workspace files need updating).
Validate: Does it handle all Stage 1 examples? Any contradictions? Edge cases?
Iterate until the user is satisfied. This is where design problems surface cheaply.
Default: same-session. Work through the spec with user review at each step. Reserve subagent handoff for complex script subcomponents only — SKILL.md and integration logic stay in the main session.
If extracting: update workspace files, clean up old locations, verify standalone operation.
The frontmatter determines discoverability and gating. Format follows the AgentSkills spec with OpenClaw extensions.
--- name: my-skill description: [description optimised for discovery — see below] homepage: https://github.com/user/repo # optional metadata: {"openclaw":{"emoji":"🔧","requires":{"bins":["tool"],"env":["API_KEY"]},"primaryEnv":"API_KEY","install":[...]}} ---
Critical:
metadataThe description determines whether the skill gets loaded. Include:
Test: would the agent select this skill for each of your Stage 1 example phrases?
| Key | Purpose |
|---|---|
| Skill identifier (required) |
| Discovery text (required) |
| URL for docs/repo |
| |
| |
| |
| Tool name for direct dispatch |
| |
OpenClaw filters skills at load time using
metadata.openclaw| Field | Effect |
|---|---|
| Skip all gates, always load |
| Display in macOS Skills UI |
| Platform filter ( |
| All must exist on PATH |
| At least one must exist |
| Env var must exist or be in config |
| Config paths must be truthy |
| Maps to |
| Installer specs for auto-setup (brew/node/go/uv/download) |
Sandbox note:
requires.binsEach eligible skill adds ~97 chars + name + description + location path to the system prompt. Keep descriptions informative but not bloated — every character costs tokens on every turn.
"install": [ {"id": "brew", "kind": "brew", "formula": "tap/tool", "bins": ["tool"], "label": "Install via brew"}, {"id": "npm", "kind": "node", "package": "tool", "bins": ["tool"]}, {"id": "uv", "kind": "uv", "package": "tool", "bins": ["tool"]}, {"id": "go", "kind": "go", "package": "github.com/user/tool@latest", "bins": ["tool"]}, {"id": "dl", "kind": "download", "url": "https://...", "archive": "tar.gz"} ]
| Token | Meaning |
|---|---|
| This skill's directory (OpenClaw resolves at runtime) |
| Agent's workspace root |
{baseDir}<workspace>/{baseDir}/patterns/No automatic installation available. Please visit the source repository for installation instructions.
View Installation Instructions
1,500+ AI skills, agents & workflows. Install in 30 seconds. Part of the Torly.ai family.
© 2026 Torly.ai. All rights reserved.