Most people use AI in chat mode: new thread, same explanation, context gone. Agents keep USER.md, run on a schedule, and follow playbooks you define once. Skills are those playbooks.
You package a workflow (morning briefing, inbox triage, Slack formatting) into a folder the agent discovers. Bundled skills ship with OpenClaw. You can add workspace skills, install from ClawHub, or write your own.
The hard part is not adding skills. It is knowing which ones are worth the folder. We are not cataloging every ClawHub listing. This is a map of types that work, habits that hold up after a week of real use, and mistakes that cost an afternoon.
What a skill actually is
People say skills are "just markdown files." They have the file right. They miss the folder.
A skill is a directory with SKILL.md (YAML frontmatter plus instructions). Often there is more: scripts/, references/, assets/. OpenClaw loads from workspace paths, ~/.openclaw/skills, bundled skills, and others. metadata.openclaw can hide skills when required binaries or env vars are missing.
In human terms: SOUL.md and USER.md are who the agent is and who you are. A skill is a recipe card in the drawer: pull it when that dish comes up, not the whole cookbook every night.
Six skill types worth building
Good skills pick one job. Messy ones try to be standup helper, deploy runbook, and style guide in the same folder.
Workflow automation
Standup drafts, weekly recaps, meal-plan refresh, ticket formatting. Pair with crons; see 5 OpenClaw Automation Workflows.
In human terms: You stopped typing the same Monday email. The agent sends the draft; you edit one line and ship it.
Tool and integration runbooks
Your stack's footguns: which flag works, which endpoint to call. See web search and AgentMail quickstarts.
Verification
Checklists before the agent says "done": confirm the cron message landed, read back key fields.
Memory and context helpers
Brain-dump interviews that write USER.md. See What OpenClaw Setup Really Takes.
Data fetch and summarize
Dashboard → summary → chat. Keep credentials in env, not pasted in Telegram.
Quality and safety guardrails
Confirm before external send, no secrets in output. Random ClawHub installs are a real risk; see security guide.
How to write a skill that gets used
Don't restate what the base model already knows. Put in your conventions and footguns. Build a Gotchas section over time; that's what you read when something breaks.
In human terms: The description is the label on the drawer. "Spices" never gets opened when you need cumin. Write for trigger matching: "Split and merge PDFs with local CLI when user mentions PDF editing" beats "helpful document skill."
Give constraints, not a 47-step script. Store setup in config.json. Scripts beat prose for brittle steps.
Install, build, or share
| Situation | Move |
|---|---|
| Common task | Bundled skill or official quickstart |
| Personal workflow | Workspace skill under /skills |
| Community skill | Read SKILL.md and scripts first |
Run openclaw skills list --eligible. Skills work on Easy and Premium; voice-heavy skills may need Premium (voice notes are Premium-only on InstantClaw).
Three starters to try
- Morning context: draft briefing from prefs when user asks for daily summary.
- Gotcha capture: append to references/gotchas.md when user says "remember this gotcha."
- Pre-send check: three-item checklist before posting to Slack or Telegram.
Reference: OpenClaw skills docs and the bundled skill-creator in the OpenClaw repo.
Mistakes we see often
- Skill where SOUL.md or USER.md would fit better
- Description too vague to trigger
- 500-line SKILL.md the agent half-ignores
- ClawHub install without reading scripts
- Voice assumptions on Easy (text-only tier)
How skills work on InstantClaw
- Teach in chat: Skill files live in your workspace on a dedicated VM. Most users never SSH in to drop folders manually.
- Easy and Premium: Skills work on both tiers. Voice note processing is Premium-only; Easy is text on Telegram and Slack.
- Trust matters: We are shipping a verified plugin store. Until then, treat third-party skills like third-party plugins: read before you install.
The bottom line
Self-hosters wire skills into VPS configs, watch load order, and patch when upgrades break paths. InstantClaw users teach the agent; we run the infrastructure and roll out OpenClaw updates.
Pick one task. Write a small skill. Run it a week. Add to Gotchas when it fails. That beats a skill library you never finish.
Frequently asked questions
- What is an OpenClaw skill?
- A skill is a directory with a SKILL.md file: YAML frontmatter (name, description) plus instructions. It can include scripts, references, and assets. OpenClaw loads bundled skills, workspace skills, and skills from paths like ~/.openclaw/skills.
- Where do OpenClaw skills live?
- Workspace skills typically live under /skills in your OpenClaw workspace. Managed local skills use ~/.openclaw/skills. Bundled skills ship with the install. See docs.openclaw.ai/tools/skills for load order and precedence.
- Should I install ClawHub skills or write my own?
- Start with bundled skills and official quickstarts for common tasks. Write your own for personal workflows (standup format, your stack's footguns). Read any third-party SKILL.md and scripts before installing; quality varies.
- What is the difference between skills and SOUL.md or USER.md?
- SOUL.md and USER.md shape who the agent is and who you are. Skills teach repeatable tasks: how to run a workflow, use a tool, or verify output. Personality files are always on; skills trigger when the request matches the description.
- Do I need Premium to use skills on InstantClaw?
- No. Skills work on Easy and Premium. Voice or multimodal skills may assume voice input; on InstantClaw, voice note processing is Premium-only. Easy is text on Telegram and Slack.
Related reading
Want the agent without the server work?
Deploy OpenClaw in under a minute. Skills live in your workspace; we run the infrastructure.