Source Code

Skillcraft — OpenClaw Skill Designer

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

Model Notes

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:

Read the pattern files in {baseDir}/patterns/ more carefully before architecting
Spend more time on Stage 2 (capability discovery) — enumerate OpenClaw features explicitly
Be more methodical in Stage 4 (spec) — write out the full structure before implementing
Consult https://docs.openclaw.ai when unsure about any OpenClaw feature

The Design Sequence

Stage 0: Inventory (Extraction Only)

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.

Stage 1: Problem Understanding

Work through with the user:

What does this skill do? (one sentence)
When should it load? Example phrases, mid-task triggers, scheduled triggers
What does success look like? Concrete outcomes per example

Stage 2: Capability Discovery

Generalisability

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

Skill Synergy (Particular Only)

Scan <available_skills> from the system prompt for complementary capabilities. Read promising skills to understand composition opportunities.

OpenClaw Features

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	`/concepts/messages`
Cron/scheduling	`/automation/cron-jobs`
Subagents	`/tools/subagents`
Browser	`/tools/browser`
Canvas UI	`/tools/` (canvas)
Node devices	`/nodes/`
Slash commands	`/tools/slash-commands`

See {baseDir}/patterns/composable-examples.md for inspiration on combining these.

Stage 3: Architecture

Based on Stages 1–2, identify which patterns apply:

If the skill...	Pattern
Wraps a CLI tool	`{baseDir}/patterns/cli-wrapper.md`
Wraps a web API	`{baseDir}/patterns/api-wrapper.md`
Monitors and notifies	`{baseDir}/patterns/monitor.md`

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.

Stage 4: Design Specification

Present proposed architecture for user review:

Skill structure — files and directories
SKILL.md outline — sections and key content
Components — scripts, modules, wrappers
State — stateless, session-stateful, or persistent (and where it lives)
OpenClaw integration — which features, how they interact
Secrets — env vars, keychain, config file (document in setup section, never hardcode)

State locations:

<workspace>/memory/ — user-facing context
{baseDir}/state.json — skill-internal state (travels with skill)
<workspace>/state/<skill>.json — skill state in common workspace area

If 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.

Stage 5: Implementation

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.

Create skill directory + SKILL.md skeleton (frontmatter + sections)
Scripts (if any) — get them working and tested
SKILL.md body — complete instructions
Test against Stage 1 examples

If extracting: update workspace files, clean up old locations, verify standalone operation.

Crafting the Frontmatter

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: metadata must be a single-line JSON object (parser limitation).

Description — Write for Discovery

The description determines whether the skill gets loaded. Include:

Core capability — what it does
Trigger keywords — terms users would say
Contexts — situations where it applies

Test: would the agent select this skill for each of your Stage 1 example phrases?

Frontmatter Keys

Key	Purpose
`name`	Skill identifier (required)
`description`	Discovery text (required)
`homepage`	URL for docs/repo
`user-invocable`	`true`/`false` — expose as slash command (default: true)
`disable-model-invocation`	`true`/`false` — exclude from model prompt (default: false)
`command-dispatch`	`tool` — bypass model, dispatch directly to a tool
`command-tool`	Tool name for direct dispatch
`command-arg-mode`	`raw` — forward raw args to tool

Metadata Gating

OpenClaw filters skills at load time using metadata.openclaw:

Field	Effect
`always: true`	Skip all gates, always load
`emoji`	Display in macOS Skills UI
`os`	Platform filter (`darwin`, `linux`, `win32`)
`requires.bins`	All must exist on PATH
`requires.anyBins`	At least one must exist
`requires.env`	Env var must exist or be in config
`requires.config`	Config paths must be truthy
`primaryEnv`	Maps to `skills.entries.<name>.apiKey`
`install`	Installer specs for auto-setup (brew/node/go/uv/download)

Sandbox note: requires.bins checks the host at load time. If sandboxed, the binary must also exist inside the container.

Token Budget

Each 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 Specs

"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"}
]

Path Conventions

Token	Meaning
`{baseDir}`	This skill's directory (OpenClaw resolves at runtime)
`<workspace>/`	Agent's workspace root

Use {baseDir} for skill-internal references (scripts, state, patterns)
Use <workspace>/ for workspace files (TOOLS.md, memory/, etc.)
Never hardcode absolute paths — workspaces are portable
For subagent scenarios, include path context in the task description (sandbox mounts differ)

References

Pattern files: {baseDir}/patterns/ (cli-wrapper, api-wrapper, monitor, composable-examples)
OpenClaw docs: https://docs.openclaw.ai/tools/skills
ClawHub: https://clawhub.com

skillcraft