SkillHub Publish Auditor#

Overview#

Audit an agent Skill as a public package, not a private note. The goal is to catch release blockers, weak discovery metadata, unsafe scripts, credential leaks, unfinished template markers, and packaging issues before upload.

Quick Start#

Run the bundled auditor from the skill folder you want to publish:

bash

python <this-skill>/scripts/audit_skill_release.py <skill-folder> --markdown

Use --json for CI or automated marketplace checks.

Release Standard#

Apply four gates in order:

1. Structure: SKILL.md exists, frontmatter parses, name is publishable, and no unfinished template markers remain.
2. Discovery: description starts with Use when, names concrete triggers, and does not summarize the whole workflow.
3. Safety: scripts and references do not read secrets, hide instruction-override payloads, run destructive shell commands, or send local data out.
4. Packaging: resource links resolve, examples are intentional, auxiliary files are not clutter, and UI metadata is useful.

Read references/release-checklist.md when judging borderline findings or preparing a manual review.

Required Behavior#

• Treat every file in the candidate skill as publishable supply-chain content.
• Inspect scripts before recommending release, even when SKILL.md looks clean.
• Mark issues as blocker, warn, or info; keep each finding specific.
• Recommend exact file-level fixes, but do not auto-edit unless the user asks.
• If the skill executes code, require at least one local smoke test or explain why no test can run.

Common Mistakes#

MistakeFixPublishing template textRemove unfinished markers and generated instructions.Description explains workflowRewrite it as trigger conditions only.Security review stops at SKILL.mdScan every script, reference, config, and example.Marketplace metadata ignoredCheck agents/openai.yaml when present."Harmless" helper scripts read env filesFlag environment, SSH, credential, token, and key access.

Output Contract#

End with a release decision:

• READY: no blockers.
• READY_WITH_WARNINGS: warnings remain, but no release blocker.
• BLOCKED: one or more blockers must be fixed before publishing.