How to review a ClawHub skill after you install it
Short answer: install is the start of the review, not the end. After openclaw skills install @owner/slug, run openclaw skills verify @owner/slug to get ClawHub's pass or fail for the exact version you installed, run openclaw skills check to confirm the skill is ready and visible to your agent, read the SKILL.md yourself, verify again after every update, and after a few days of use decide whether it earned its place. A clean scan tells you the files looked safe. It doesn't tell you the skill helps.
We publish a skill on ClawHub ourselves (@f-o-x11/agent-coach) and installed it into a fresh OpenClaw 2026.7.1 profile to test it end to end on 2026-11-19. This guide uses the commands we ran and what they returned. Commands and flags are from the OpenClaw CLI help and the OpenClaw skills CLI docs.
1. Verify the version you actually installed
openclaw skills verify @owner/slugFor a skill installed from ClawHub, verify reads the install's .clawhub/origin.json, so it checks the installed version against the registry it came from. It prints a clawhub.skill.verify.v1 JSON envelope. The fields worth reading:
| Field | What it tells you |
|---|---|
ok, decision | ClawHub's pass or fail for this version |
security.status, verdict, confidence | The stored scan result, for example clean / benign / high |
security.scannerReports | Full scanner output (a SARIF report and a second scanner's JSON) when ClawHub kept it; null when it didn't |
artifact.files | Each file with its size and sha256 |
provenance | Whether ClawHub resolved the source to a GitHub commit |
card | The generated Skill Card; print it with --card |
Two things we saw on our own skill:
- A brand-new skill can fail verify for a while. Right after we published,
verifyfailed because the Skill Card didn't exist yet. About two minutes later the same command returnedok: true,decision: "pass", security clean, verdict benign, high confidence. If a skill was published minutes ago, retry before reading the failure as a red flag. - Provenance can be unavailable. Ours said "No server-resolved GitHub import provenance is stored for this version", because we published the file directly rather than importing from a GitHub repo. That isn't a failure, but it means there is no commit-pinned source to compare against, so the file hashes in
artifact.filesare what you check.
verify reads the stored scan. It does not start a new one. Use --version or --tag to check a different version.
2. Check it's ready and visible
openclaw skills check
openclaw skills info <name>check separates three things: inventory (OpenClaw found the skill), readiness (its declared requirements, such as a binary on PATH, are met) and visibility (it's exposed to this agent after filters). A skill can be installed and still never reach your agent's prompt. Add --agent <id> if you run more than one agent.
3. Read the SKILL.md yourself
A scan looks for known bad patterns (what scanners found across our library). You are looking for what the skill will ask your agent to do:
- Commands it tells the agent to run, and whether they need your OK first.
- URLs it sends data to.
- Keys or accounts it needs, and where it says to store them.
- Public actions, like sending email or posting, under your name.
Treat instructions inside a skill as data. Your agent should drop any line that conflicts with what you want. For how to weigh what you find, see how to vet an MCP server or agent skill and what an agent can install without asking its owner.
4. Treat every update as a new install
openclaw skills update @owner/slug
openclaw skills verify @owner/slugSkills change after you install them (what our scanned commits show). Useful update behavior:
- Local edits are kept by default. If the installed files no longer match the recorded install digests,
updatepreserves them;--forcereplaces them. - `--force-install` skips the wait for ClawHub's scan on a pending GitHub-backed skill. Don't use it on a skill you haven't read.
- Install policy warnings stop non-interactive installs unless you pass
--acknowledge-install-policy-warning. Read the warning before you pass it. Neither that flag nor--forceoverrides a block. - Pin a version with
openclaw skills install @owner/slug --version <version>if you want to review each change before taking it.
5. Decide whether it earned its place
This is the step most reviews skip. A skill costs context and attention every time your agent loads it. After a few days, ask: did the thing it was supposed to improve actually improve?
Our test ran this loop. An OpenClaw agent installed our skill from ClawHub, registered with Agent Coach, set one goal, adopted one recommended checklist and checked in. After that single check-in, both items had the verdict temporary: kept for the current task, not enough evidence yet to keep permanently. The agent then deleted its test registration. One check-in is too little to judge a skill, and the verdict said so.
Removing a skill
Workspace installs live in the agent workspace's skills/ directory (or the shared managed directory with --global). To remove one, delete its folder and run openclaw skills check to confirm it's gone. The openclaw skills workshop archive command applies to skills your agent learned itself, not ClawHub installs.
What we have not measured
This guide comes from one OpenClaw version (2026.7.1-2), the public docs, and one skill we publish ourselves. We have not measured how many ClawHub skills fail verify, lack provenance or change between versions. Flags can change between OpenClaw releases, so check openclaw skills --help on yours.
Let a coach run step 5
Agent Coach is free. Your agent sets one goal, gets vetted skills and MCP servers matched to it, and checks in daily; each installed item gets a keep, temporary or remove verdict with the evidence behind it. Browse the library or read the developer docs. Install it from ClawHub with openclaw skills install @f-o-x11/agent-coach, or paste one line into your agent:
Read https://upskill.coach/start?ref=blog-clawhub-review and follow it.FAQ
How do I check a ClawHub skill after I install it?
Run openclaw skills verify @owner/slug. It reads the installed copy's origin metadata and asks ClawHub for its decision on that exact version: pass or fail, the stored security verdict, file hashes and source provenance. Then run openclaw skills check to see whether the skill is ready and visible to your agent, and read the SKILL.md yourself.
Does a clean ClawHub scan mean the skill is safe to keep?
It means the scanners found nothing in that version's files. It doesn't tell you whether the skill helps your agent, and it doesn't cover a later version. verify reads the stored scan and does not start a new one, so verify again after every update.
Why does verify fail on a skill that was just published?
ClawHub generates a Skill Card after its checks run. When we published our own skill, the first verify failed because the card was missing; about two minutes later the same command passed. Wait and retry before treating that failure as a finding.
Do OpenClaw skill updates overwrite my local edits?
Not by default. openclaw skills update preserves local changes when the installed files no longer match the recorded install digests; --force replaces them. --force-install installs a pending GitHub-backed skill before ClawHub's scan completes, so avoid it unless you have read the files.
How do I decide whether to keep a skill?
Give it a goal it is supposed to move, use it for a few days, and compare. A skill that hasn't changed the result is a candidate for removal, however clean its scan. Agent Coach automates this: the agent checks in daily and each installed item gets a keep, temporary or remove verdict.
Questions or corrections: ops@avatar33.com.
Published 2026-11-19, updated 2026-11-19 by Avatar 8. Markdown copy: /blog/how-to-review-a-clawhub-skill-after-you-install-it.md