skill-author

Author, harden, and publish agent skills for ClawHub — write skills that conform to ClawHub conventions (frontmatter, structure, naming, versioning, licensing), run pre-publish hardening, and remediate clawhub.ai security-audit findings. Use when creating a skill for ClawHub, preparing a new version to publish, or when given a clawhub.ai///security-audit page.

ClawHub Skill Maker

Make agent skills that are well-formed for ClawHub. Four capabilities, in workflow order:

1. Author — write a skill that conforms to ClawHub conventions (frontmatter, structure, naming, versioning, license, description style), so it passes review cleanly the first time.
2. Harden — pre-publish check: completeness, dead links, secrets, build artifacts, doc↔code consistency, security by design.
3. Remediate — fix findings from a clawhub.ai/<owner>/<slug>/security-audit page (hard findings → code/config changes; soft findings → docs/policy clarification).
4. Publish & verify — clawhub publish with a proper changelog, then poll through the moderation queue until Latest: flips.

Know the review model before you fix anything

ClawHub reviews run in two layers:

Layer	Produces	How to handle
Static pattern scan	Hard findings with file:line (command injection, hardcoded secrets, instruction hijacking, memory poisoning, remote payload fetch, insecure dependencies)	Must be fixed in code/config — wording changes do not count
LLM review	Soft findings with confidence scores (language/locale policy, agency & consent, doc/code consistency, missing risk disclosures)	High confidence (≥90%) = real; lower = verify first. Usually fixed in docs

False positives happen: path-traversal defense code gets flagged as "credential access"; reading the skill's own workflow JSON gets flagged as "memory poisoning". Never delete a guard to "fix" a finding — keep it and explain it in the report to the user.

1. Author a ClawHub-conformant skill

Follow the conventions in references/authoring-guide.md. The short version:

1. Check the slug — clawhub search "<keywords>" to avoid collisions (slug defaults to the sanitized folder name).
2. Scaffold the layout — SKILL.md, README.md, LICENSE, VERSION, references/ (see guide §Structure).
3. Write the frontmatter to the field reference — name, description (triggers first), version, license, homepage, compatibility, allowed-tools, metadata.openclaw.emoji (guide §Frontmatter).
4. Keep SKILL.md a short procedure — the shortest steps that reproduce the result (<10k characters); long reference, examples, and templates move to references/ files pointed to from the step that needs them.
5. Write the description as triggers — "Use when …" situations first, then what the skill produces. The description is what search and skill-matching match against.
6. Apply security by design (section 3 below) while writing, not after review.
7. Add README + LICENSE + VERSION — the README is for humans (capabilities, structure, install); LICENSE is the declared license; VERSION mirrors the frontmatter version.

2. Harden before publishing

Run the checklist in references/publish-checklist.md. The ten points:

1. Diff local vs registry version (new releases must be a deliberate superset)
2. Frontmatter complete (field reference in the authoring guide)
3. No dead links — every references/*.md cited in docs must exist
4. Doc↔code consistency — every subcommand/env var/file named in docs exists in code
5. No build/test artifacts (__pycache__/, test_output/) in the package
6. No hardcoded private IPs, secrets, or personal data
7. LICENSE present and matches frontmatter license
8. VERSION matches frontmatter version
9. Version bumped (semver) with a changelog that lists the changes
10. Logged in (clawhub whoami)

3. Security by design (authoring-time self-check)

Class	Question while writing	Fix
Command injection	Do shell examples interpolate user-controlled values (titles, paths, filenames)?	Allowlist sanitize + quote every expansion + verify destination stays inside the target dir; prefer file tools over constructed shell commands
Data egress	Does any bundled config point at a non-loopback default endpoint? What data leaves the machine?	Default to loopback-only; print a disclosure at runtime for non-local endpoints; document exactly what is transmitted and where outputs are written
Supply chain	Do install instructions clone mutable branches without pins or locks?	Pin reviewed commits (git checkout --detach <sha>), lock dependencies, verify repos still exist via the GitHub API (renamed/migrated repos 404), recommend venv + non-root

Full pattern → fix mapping with real examples: references/audit-patterns.md.

4. Remediate a security audit

When the user shares a security-audit page (fetch it with web_fetch):

1. Split findings — hard (file:line, fix in code/config) vs soft (confidence + suggestion, usually docs/policy).
2. Fix hard findings one by one — locate, apply the matching pattern from references/audit-patterns.md, verify (compile/parse the file, run the skill's own test entry, grep for leftover old values).
3. Judge soft findings — real contradiction/risk → fix (add confirmations, disclosures, scope limits, language notes; remove phantom references); false positive → keep the code and note it.
4. While you're in there: dead links, artifacts, LICENSE/VERSION, version bump, changelog.
5. Publish and verify (below).
6. Report to the user — finding → fix table for every hard finding, key soft findings, and a false-positives section.

5. Publish and verify

bash

cd <parent of the skill folder>
clawhub whoami                                        # confirm logged in
clawhub publish ./<slug> --version X.Y.Z --changelog "<itemized changes>"
# success: OK. Published <slug>@X.Y.Z (<versionId>)
# then poll through the pending.publication moderation queue:
clawhub inspect <slug> 2>&1 | grep "Latest:"          # wait for "Latest: X.Y.Z"

Registry notes: the listing page always shows License: MIT-0 (the platform's fixed registration term) — the in-package LICENSE file is the declared license installers actually get; a mismatch is not an error. The registry auto-generates a skill-card.md; never add one to source. Full commands and polling loop: references/publish-checklist.md.

References

• references/authoring-guide.md — ClawHub skill-authoring conventions: frontmatter field reference, slug/naming rules, versioning, tags, file layout, license model, description writing
• references/publish-checklist.md — completeness / publish / verification checklist with exact CLI commands (registry diff, moderation polling, repo liveness checks)
• references/audit-patterns.md — audit pattern → fix mapping (command injection, data egress, supply chain, locale, agency/consent, doc/code consistency, false positives) from two real remediation cases