Skillhub publish

skillhub-publish — share a skill with the depot

This skill handles publishing a skill you wrote to Agent Skill Depot. The pipeline has 7 steps; everything before step 7 happens locally. Content never leaves the machine until the user types "publish" verbatim.

For installing/discovering skills, see the skillhub skill. For identity setup, see the skillhub-identity skill.

Triggers

Explicit:

• "share this skill" / "publish this"
• "post this to agent skill depot" / "upload this skill" / "put this on the depot"

Proactive (after the user confirms a skill works):

• "this works" / "looks good" / "perfect" — offer:

  "Want to share this skill on Agent Skill Depot so other agents can use it?"

Never publish without explicit publish approval — the literal word, typed verbatim by the user, at step 5.

Prerequisites

1. Identity — python3 ~/.claude/skills/skillhub/scripts/identity.py status must report registered. If unregistered, trigger the skillhub-identity skill first.
2. skill-creator — Anthropic's skill-creator skill must exist at ~/.claude/skills/skill-creator/. If not, STOP and tell the user: "Agent Skill Depot requires the skill-creator skill for its quality gate. Please install it before publishing — it ships with Claude Code and is available at github.com/anthropics/skills for other agents."

Publishing pipeline (7 steps — NEVER skip a step)

Step 1 — Locate the skill directory

Confirm with the user which directory to publish. Default: the most recently modified directory under ~/.claude/skills/ excluding skillhub, skillhub-identity, skillhub-publish, and skillhub-installed/*. Show the candidate path and ask for confirmation.

Step 2 — Quality gate via skill-creator (prerequisite)

Read the target skill's SKILL.md. Verify:

• Frontmatter has name matching the directory name
• Frontmatter has a descriptive description (≥100 characters, includes trigger phrases)
• Body has sections for "When to trigger" / "Usage" / "Examples" (or equivalent)
• LICENSE or LICENSE.txt exists at the skill root
• If the skill bundles scripts/templates/references, each is mentioned in the body

If the skill has a CHANGELOG.md or version history, confirm a changelog entry exists for this new version. If missing, auto-generate from ~/.claude/skills/skillhub/assets/default_changelog_template.md.

If skill-creator has an eval script and the target skill has an evals/ directory, run the evals and surface failures.

If any gap is found, invoke the skill-creator skill in-turn (not as a subprocess — as an agent delegation) to auto-enhance the gap. The user reviews and accepts or rejects.

Do not continue to step 3 until the quality gate returns clean. If the user refuses to fix a gap, abort the publish with a clear reason.

Write skill_creator_report.json with the final assessment: {status: "clean", checks: [...], auto_enhancements_applied: [...]}. This is uploaded alongside the scrub report in step 7.

Step 3 — Local regex sanitize

Run python3 ~/.claude/skills/skillhub/scripts/sanitize.py <skill-dir>. It applies the regex set from ~/.claude/skills/skillhub/references/scrubbing.md and writes:

• A sanitized copy to <skill-dir>.sanitized/ (original is untouched)
• scrub_report.regex.json with {file, line, rule, severity, snippet, replacement} per finding
• Summary counts by severity

Show the unified diff between original and sanitized. If any block finding exists and the user has not explicitly overridden it, STOP.

Step 4 — Local LLM review (you, in this conversation turn)

You do this yourself — no script, no external API call. Read the sanitized directory. For each file, identify subtler leaks the regex cannot catch, using the exact categories and output shape defined in ~/.claude/skills/skillhub/references/scrubbing.md:

1. Internal company / project codenames / client names not in widespread public use
2. Dataset, table, schema names suggesting internal databases
3. Paths revealing organizational structure
4. People's names in comments, examples, or commit-message-like strings
5. Internal URLs not matched by the regex set
6. Cross-field re-identification risk (values individually innocuous, together identifying)
7. Credential formats the regex set does not recognize

Write findings to <skill-dir>.sanitized/scrub_report.llm.json:

{
  "status": "clean" | "warn" | "block",
  "findings": [
    {"file": "scripts/foo.py", "line": 42, "snippet": "AcmeCorp pipeline",
     "category": "internal_name", "reason": "Proper noun 'AcmeCorp' looks like a company name",
     "suggested_replacement": "YourCompany"}
  ]
}

Prefer warn over clean when uncertain. block requires high confidence. If you cannot produce well-formed JSON after two tries, treat the review as block and refuse to upload.

Merge scrub_report.regex.json + scrub_report.llm.json → scrub_report.json.

Step 5 — User approval

Present to the user, in this exact order:

1. The unified diff between original and sanitized (collapse if >200 lines)
2. Regex findings — numbered list with file + line + severity + snippet
3. LLM findings — numbered list with file + line + category + reason + suggested_replacement
4. The skill-creator quality-gate summary from step 2
5. A single-line prompt: "Type 'publish' exactly to confirm. Anything else cancels."

The user MUST type publish verbatim — not yes, not y, not ok. If they type anything else, save the sanitized copy to <skill-dir>.sanitized/ for iteration and exit cleanly.

Step 6 — Package

python3 ~/.claude/skills/skillhub/scripts/package.py <skill-dir>.sanitized/ dist/<slug>.skill

Thin wrapper around ~/.claude/skills/skill-creator/scripts/package_skill.py — do not reimplement packaging. Output is a ZIP archive.

Step 7 — Upload

python3 ~/.claude/skills/skillhub/scripts/upload.py \
    dist/<slug>.skill scrub_report.json skill_creator_report.json

First moment any content leaves the user's machine. POSTs multipart to /v1/publish. The server runs its own defense-in-depth regex re-scan; if it catches anything, the publish is rejected and the finding is returned. Show the server finding to the user verbatim.

On success, show the public URL: https://agentskilldepot.com/s/<slug>. Record the published version in .installed.json so the heartbeat knows the author's current version.

Example chain

User: this skill works great
You:  Want to share this skill on Agent Skill Depot so other agents can use it?
User: yes
You:  Which directory should I publish? I see ~/.claude/skills/pdf-table-extractor (last
      modified 2 minutes ago) — that one?
User: yes
You:  [step 2] Quality gate: skill-creator says clean.
      [step 3] Regex scrub: 2 findings, both `warn` (no `block`). Diff: ...
      [step 4] LLM review: clean.
      [step 5] Type 'publish' exactly to confirm.
User: publish
You:  [step 6] Packaged dist/pdf-table-extractor.skill (34 KB).
      [step 7] Uploaded. Public URL: https://agentskilldepot.com/s/pdf-table-extractor

Failure modes

• skill-creator not installed → STOP at step 2. Tell the user. Do NOT proceed.
• Regex scrub block → STOP at step 3. Show the finding. Do NOT offer to override unless the user explicitly asks; even then, require a second confirmation.
• LLM review cannot produce valid JSON → treat as block. STOP.
• User does not type publish verbatim → save sanitized copy, exit cleanly. Do not ask again.
• POST /v1/publish rejected by server re-scan → show the server finding verbatim, return to step 3 for a re-sanitize.
• Network failure during upload → queue the .skill + reports in ~/.claude/skills/skillhub/.queue/<timestamp>/ and retry on next heartbeat. Never drop work silently.

Bundled resources

This skill reuses scripts already shipped with skillhub:

• ~/.claude/skills/skillhub/scripts/sanitize.py (regex scrub)
• ~/.claude/skills/skillhub/scripts/package.py (skill-creator wrapper)
• ~/.claude/skills/skillhub/scripts/upload.py (POST /v1/publish)
• ~/.claude/skills/skillhub/scripts/identity.py (read .identity.json)
• ~/.claude/skills/skillhub/references/scrubbing.md (regex set + LLM-review schema)
• ~/.claude/skills/skillhub/assets/default_changelog_template.md