How to use the Skill Release Gate agent skill

What is an Agent Skill and what does this skill make the agent do?

An Agent Skill is a folder that gives an AI agent extra instructions and files for a specific task. Its core file is SKILL.md, and it can also include evals/, assets/, references/ and scripts/ folders.

Skill Release Gate tells the agent to evaluate such a bundle before it is published or distributed. The agent reads the bundle's evaluation files, runs a bundled Python script that checks structure, trigger accuracy, artifact comparisons, safety findings, file-hash integrity and target-host compatibility, and then reports the results layer by layer.

• Structural integrity: directory name matches frontmatter name, required metadata exists, and no unreferenced companion file ships.
• Trigger quality: positive prompts activate the skill and near-miss prompts stay inactive, with precision and recall reported.
• Artifact comparison: the same assertions are applied to baseline and with-skill output.
• Script and safety evidence, installed-tree hash verification, and a target-host portability matrix.

Install the skill from the ZIP or the Claude apps

Download the ZIP for skill-release-gate from this page. You can install it in one of two ways.

For Claude Code, unzip the download into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ folder to share it with that project.

In the Claude apps, upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project

Trigger the skill with natural requests

You do not need a special command. Ask in plain language that you want a release review of a skill bundle. The agent should load the skill when your request matches its description, which covers evaluating a bundle before release.

Be specific about the candidate folder so the agent can resolve the absolute path to the bundle under review.

• "Use the release gate to check the skill in ./my-skill before I publish it."
• "Review this Agent Skill bundle for structural problems and trigger accuracy."
• "Run the pre-release evaluation on this skill folder and show me the portability matrix."

Run the evaluator and read the report

The skill's workflow asks the agent to show the exact resolved command before running it. The installed evaluator is scripts/evaluate_skill.py under the skill's root folder. For the shipped teaching fixture, the argv is built from python3, the absolute path to that evaluator, --fixture-demo, and the absolute path to the target bundle.

The report must include checksPassed, fixturePassed, localEvidenceReady, trustAnchorValid, productionReady and passed, along with the evidence root, failed checks, precision, recall, every raw trigger observation, per-case repeated-run rates, artifact comparison, script and safety evidence, installed-tree verification and the portability matrix. The resolved script path, target path, working directory, exact argv and exit code are also recorded. Any observation that could not be captured is marked unverified.

• Read the resolved script path, target path, cwd and exact argv before accepting the result.
• Check every layer. A passing aggregate does not hide a routing, artifact, script, safety, installed-tree or portability failure.
• Treat passed as true only when productionReady also has a valid out-of-bundle trust anchor.

Understand the fixture run and production mode

The shipped run is a deterministic teaching fixture. It helps you see how the layers fit together. A fixture pass only sets fixturePassed and never makes a release claim.

Production mode is different. It requires you to replace predictions, artifacts, evidence and host capabilities with captured results, set all four captured modes, and bind each input to non-empty sources with matching SHA-256 digests. You also need an external JSON attestation whose evidenceRoot matches the report, plus the SHA-256 of that file's exact bytes from a separate trusted channel. The attestation must be a regular file outside the target bundle. The production invocation uses --attestation and --trusted-attestation-sha256 with the absolute target path, and omits --fixture-demo.

• Local checks can set localEvidenceReady, but recomputable hashes do not prove that results were really captured.
• Without an external attestation and a separate trust channel, the result cannot support a production release decision.

Safety notes before you run the script

The skill runs scripts and does not use the network, according to its listed details. The bundled Python script is executed locally against the bundle under review, so review untrusted bundles on a machine without sensitive files.

The skill also stops with a nonzero result on invalid configuration, missing or mismatched provenance, a missing or invalid trusted attestation, a file hash mismatch, a missing required capability, or a failed production gate. It never publishes, installs elsewhere, repairs evidence, creates the trust decision or weakens a threshold automatically.

• Requires Python 3 and filesystem and shell access for the agent.
• Scripts run: yes. Network use: no.
• Do not treat a bundle as ready just because SKILL.md parses or one positive prompt activates.
• Do not label a package portable if a target drops required companion files or ignores required runtime extensions.

Tips for a useful review

Keep the candidate bundle in its own folder so the evaluator checks exactly what you plan to ship. Ask for the full JSON report rather than a summary, so every layer stays visible.

Prepare the evaluation files before you start: positive and near-miss trigger cases in evals/cases.json, baseline and with-skill assertions in evals/artifacts.json, and script and safety results in evals/evidence.json. Read references/eval-contract.md to understand what each surface requires.

• Read references/eval-contract.md first; it defines the required surfaces.
• Keep the attestation file outside the bundle.
• Record unverified observations as unverified rather than filling them in.