Evaluate & score

A .skill can be run and measured, not just structurally validated. This is the native counterpart to test-prompt-and-grade authoring loops — sealed into the package instead of living only in a local workspace.

Prompt your agent

Run the eval cases declared in this workspace's contract:

skill eval . --attach

Grade only what you can honestly check — leave anything you're not sure
about as pending_human, don't claim it passed. Then compile so the
benchmark gets sealed into the package:

skill compile -m "eval" --approve

What skill eval actually checks

contract.evals[] is an optional array of test-prompt-plus-assertion cases. Grading is honest about what it can verify automatically:

• An assertion prefixed contains: "phrase", not_contains: "phrase", or regex: pattern is graded against a response you supply.
• Everything else — check: "human", or a runtime assertion with no recognized directive — is reported as pending_human, never a fabricated pass.
• Each case's executability (did the workflow itself structurally dry-run) is recorded independently of whether its assertions passed.

skill eval never calls a model itself. The agent that already ran the prompt supplies the response for grading — this command's job is structural dry-run, grading, and sealing, not inference. It never auto-mints, either: running an eval has no effect on trust state.

Sealing the result

skill eval --attach writes .skill/benchmark.json in the workspace. The next skill compile picks it up automatically and seals it into provenance/benchmark.json — no separate flag on compile itself needed.

Turning evidence into a score

Score ./file.skill:

skill score --profile release

If provenance/benchmark.json is missing or thin, tell me that means the
confidence is low — not that the quality is low. Those are different claims.

skill score maps a package's provenance/benchmark.json (plus its structural completeness and provenance-integrity digests) into @skillerr/skill-score's evidence-receipt format and prints a score, confidence, and per-dimension breakdown. Confidence and quality are tracked separately on purpose — a skill with no eval evidence at all gets a neutral quality estimate and low confidence, never scored as if it were bad.

If @skillerr/skill-score isn't installed, skill score writes the mapped assessment.json instead and tells you how to score it separately — it never silently fails or fabricates a number.

Related

• Convert a SKILL.md — evals/evals.json from a skill-creator folder maps directly into contract.evals
• Trust and security