ClawHub Publish Flow
Use this skill to publish a local skill folder to ClawHub through the already-authenticated local ClawHub account.
Core rule
Only publish skills that have already passed local review. Do not publish directly from vague drafts, half-finished folders, or unreviewed third-party skills.
Preferred workflow
- Confirm the local skill folder path.
- Check login state with
clawhub whoami. - Inspect the current remote skill with
clawhub inspect <slug>when updating an existing skill. - Decide the target version and changelog.
- Perform a public-release sensitive-data review on the exact outbound payload.
- Package the skill with
package_skill.pyif a distributable artifact is needed. - Upload using
scripts/publish_to_clawhub.js. - Verify with
clawhub inspect <slug>after upload. - Report the final remote version and URL.
Assistant-created skills policy
When the user has explicitly asked that assistant-created skills be uploaded to ClawHub, treat this skill as the single publish flow for both:
- one-off manual releases
- newly created local skills after review and commit
For newly created assistant-authored skills:
- Confirm the new skill folder exists under
workspace/skills/<slug>. - Confirm it has already been reviewed and committed locally.
- Inspect remote state with
clawhub inspect <slug>. - Versioning:
- if missing remotely: publish
0.1.0 - if already present: bump patch version
- if missing remotely: publish
- Use a short changelog such as
Initial release,Add runtime notes,Refine trigger description, orDocument proven workflow. - Publish and verify.
- If a registry sheet exists, update it too.
Rate-limit fallback for new skills
ClawHub may rate-limit new skill creation (for example: max 5 new skills per hour).
When a publish attempt for a new skill fails due to rate limiting:
- Do not keep hammering the endpoint.
- Record the pending publish locally.
- Schedule a delayed retry instead of requiring the user to remind you.
- Use the same verified publish flow when retrying.
Preferred local queue file:
workspace/research/clawhub-publish-queue.md
Queue fields:
- date
- slug
- local skill path
- target version
- changelog
- reason queued
- next retry note
Retry rule:
- wait until the rate-limit window has plausibly cleared
- then retry publication once
- verify with
clawhub inspect <slug>after success
Safety and release rules
- Do not publish without an explicit user request.
- Do not guess ownership, slug, or version when ambiguity matters.
- Prefer patch bumps for small documentation or rule updates.
- If the remote state is unclear, inspect first instead of overwriting blindly.
- If a third-party publisher skill promises features but lacks code, ignore it and use this verified local flow.
- Treat package-style publisher shells as untrusted until verified locally: if the folder is mostly README / metadata and lacks runnable implementation, do not route publishing through it.
- Do not publish vague drafts or half-finished skills just because they were created; assistant-created skills still need to be coherent and committed.
- ClawHub is a public-release path unless explicitly proven otherwise; use the highest sensitivity review bar.
- Before upload, inspect the exact outbound payload for secrets, tokens, passwords, cookies, local absolute paths, runtime artifacts,
.env, and machine-specific config. - If the skill package contains anything that should stay local-only, stop and fix the package before publishing.
Required local assumptions
This skill expects:
- local
clawhubCLI installed and authenticated - local ClawHub config present under the current user
- a real skill directory containing
SKILL.md - Node available for the publish script
Scripts
Use scripts/publish_to_clawhub.js for the actual API upload.
Arguments:
node scripts/publish_to_clawhub.js --skill-path <path> --slug <slug> --version <version> --changelog <text>
Optional:
--display-name <name>
--tag <tag> # repeatable
References
Read references/release-checklist.md before publishing when the release is important or public.
Output style
Report with this structure:
- skill
- local path
- prior remote version (if any)
- target version
- upload result
- verification result
- final URL
- risks or follow-ups
- whether publication succeeded immediately or was queued for delayed retry
Keep it short and operational.