Evaluate an Agent Skill bundle for structural integrity, trigger quality, artifact improvement, script correctness, safety, installed-tree integrity, and target-host portability before release.
SkillScore77Maintained
Automatic
◆QA & Test Engineering pack
Last verified 9/27/2026
What it changes
Passed 10 of 13 test cases with the skill, and 5 of 13 without it.
Without skill36%
With skill77%
Checks it passed that the model alone missed
The execution command includes `--fixture-demo`.
Trigger cases are inspected from `evals/cases.json`.
Safety results are inspected from `evals/evidence.json`.
and 9 more in the full report
10 fixed7 passed either way4 still failing1 broke
Pass-rate lift
+41 pp
Tokens
+35%
Agent turns
0%
Response time
-32%
Model gemini-3.6-flashMeasured 9/27/2026v117 of 22 cases passed with skill · 13 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 6 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 6
Prompt
An auditor wants to know where the baseline and with-skill evaluation assertions are stored in a candidate skill bundle `/workspace/my-agent-skill`. Which target relative path must the release gate inspect for these assertions?
What was checked, without and with the skill
Checked
Without skill
With skill
Baseline assertions are inspected from `evals/artifacts.json`.
Missed
Passed
case-11 · gemini-3.6-flash · 9/27/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md54 lines5 filesincludes scripts
Skill release gate
Use this skill before publishing or distributing an Agent Skill directory bundle.
Workflow
Resolve SKILL_ROOT to the absolute directory containing this installed
SKILL.md. Do not assume the process cwd is the installed bundle.
Resolve TARGET_ROOT from the original workspace working directory and
resolve the user-supplied candidate as an absolute TARGET_BUNDLE.
Read references/eval-contract.md from SKILL_ROOT.
Inspect the positive and near-miss trigger cases in
evals/cases.json under TARGET_BUNDLE.
Inspect the shared baseline and with-skill assertions in
evals/artifacts.json under TARGET_BUNDLE.
Inspect the explicit script and safety results in
evals/evidence.json under TARGET_BUNDLE.
Inspect the declared runtime capabilities in
assets/hosts.json under TARGET_BUNDLE and verify the target file hashes
against its assets/manifest.json.
For production, replace deterministic predictions, artifacts, evidence,
and host capabilities with captured results; set all four captured modes;
and bind every raw trigger observation, both artifacts, the complete
evidence set, and the non-empty host matrix to non-empty sources and
matching SHA-256 provenance digests. These local checks can set
localEvidenceReady, but locally recomputable hashes do not prove capture.
Obtain an external JSON attestation whose evidenceRoot matches the report,
plus the SHA-256 of its exact bytes from a separate trusted policy or
release channel. The attestation must be a regular file outside the target
bundle.
Before execution, show the exact resolved argv. The installed evaluator is
scripts/evaluate_skill.py under SKILL_ROOT. For the shipped lesson
fixture, build argv from python3, that absolute evaluator path,
--fixture-demo, and the absolute TARGET_BUNDLE. For production, use the
same installed script with --attestation,
--trusted-attestation-sha256, and the absolute TARGET_BUNDLE, without
--fixture-demo.
Return checksPassed, fixturePassed, localEvidenceReady,
trustAnchorValid, productionReady, and passed with the evidence root,
evaluation modes, failed checks, precision, recall, every raw trigger