Eval authoring¶

Use Anthropic skill-creator's evals.json schema: the top-level skill_name must match the skill frontmatter, and each case has an integer id, prompt, and expected_output. Optional files and expectations provide input files and verifiable success criteria.

Write expected_output in one to three sentences describing the behavior that should result, not the wording of a good reply. Keep each case focused and make every expectation check only what its prompt asks for. Split a case that asks several independent questions.

Judge substance rather than exact wording. If a case checks behavior already common without the skill, keep the instruction in SKILL.md and remove the case instead of sharpening it until a difference appears.

This repository stores these cases but does not run or grade them. Do not describe them as a local gate or claim that a change passed them.