How should developers record that a skill was tested in a specific environment?

[yuehua1213]

The repository explicitly says these skills demonstrate patterns, that behavior may differ from what is shown, and that developers should test thoroughly before relying on them for critical tasks. That is a valuable boundary between an implementation example and a verified outcome.

For reuse decisions, it could help to attach a small, source-linked evidence record to a skill version: test task, Claude surface/model, environment and tools, observed result, failures, limitations, and date. This would not claim universal quality; it would only show what was reproduced under stated conditions.

Is there a recommended format or location for maintainers and downstream users to publish that kind of skill-level test evidence?

[ryanportfolio]

I don't know of any recommended format or location. In the spec (https://agentskills.io/specification), compatibility states requirements, not verification, and metadata is the only slot for anything the spec doesn't define.

skill-creator's eval workflow is close. Test cases live in evals/evals.json inside the skill directory, and each run is graded with per-expectation pass/fail and evidence (schema in skills/skill-creator/references/schemas.md)

So I'd commit evals/evals.json, put surface/model/date/pass rate and hit limitations in a TESTING.md next to SKILL.md, and point at it with a namespaced key like yourorg.evidence: TESTING.md in metadata