Publish a skill

This guide walks the full publish path — scaffold, write, sign in, ship — in about ten minutes.

1. Install the CLI

npm install -g @amitte-ai/cli
skills --version

If skills isn't on your PATH after install, your global npm bin directory probably isn't either. Run npm config get prefix to find it.

2. Scaffold

skills init claude-code/launch-checklist -d "Pre-launch QA for shipping a Claude Code project"
cd claude-code-launch-checklist

This creates two files:

• manifest.json — the canonical metadata (id, version, capabilities, etc.)
• skill.md — the actual instruction content the agent will read

The directory name uses hyphens because filesystems don't love slashes.

3. Write the skill

Open skill.md and replace the scaffolded sections with real content.

A good skill is:

• Specific. "How to ship an iOS app" beats "iOS development".
• Bounded. Cover one task end-to-end; don't try to teach a whole framework.
• Actionable. Include the exact commands, decisions, and pitfalls — not just background.
• Honest about limits. Note what the skill is not for.

The judge LLM scores on these dimensions plus prompt-injection safety.

4. Validate locally

skills eval

This runs the same first three pipeline stages the registry will run server-side: schema validation, prompt-injection lint, artifact hash check. Catches mistakes before you spend a round trip.

5. Sign in

skills login

You'll see a one-time code and a URL — open the URL, enter the code, approve "Authorize Amitte" on GitHub. The CLI saves a token at ~/.config/amitte/credentials.json (mode 0600).

The token is a GitHub access token. Amitte uses it to identify you, never to make changes on your behalf.

6. Publish

skills publish

The CLI:

1. Re-runs local eval (unless --skip-eval).
2. Computes the artifact SHA-256.
3. POSTs { manifest, artifact_base64, publisher_handle } to /v1/submit.
4. The server runs six stages: schema → injection-lint → duplicate-check → adversarial-eval → content-eval → policy.
5. On pass, you get back a submission_id and a published URL.

✓ published
Submission: 33ef869f-62b3-4cbb-bed3-f9aa6041cb0e
URL:        https://amitte.com/s/claude-code/launch-checklist

7. (Optional) Watch the pipeline

If you publish in async mode (coming soon — currently inline), poll status:

skills status <submission_id> --watch

What "trust tier" means

Every published skill carries a tier badge:

Tier	How
unverified	Default. Anyone with a GitHub login can publish.
verified	Publisher proved control of a domain via DNS TXT. See Verify your domain.
official	Manually awarded to the platform's curated skills.

Your tier applies to future publishes. Existing skills keep the tier they had at publish time. To bring older skills up to your current tier, re-publish them with a version bump.

Common rejections

• prompt-injection-lint: artifact contains overt prompt-injection patterns — your skill.md includes "ignore previous instructions" or similar. Rephrase.
• policy: license "<x>" is not permitted — pick a license from the OSI / Creative Commons approved lists. CC-BY-4.0 is the platform default.
• duplicate: <id>@<version> already exists — bump the version in manifest.json.

That's the full loop. From here read Install a skill for the consumer side, or Verify your domain for the trust upgrade.