Add changelog evals#73
Conversation
Skill Eval Resultsdocs-fix-changelog
Overall: 46/51 expectations passed (90%) The skill instructions are well-structured and comprehensive. Two areas to note:
docs-review-changelog
Overall: 43/47 expectations passed (91%) Skill instructions are thorough and well-specified. The single recurring gap is eval 8 in both skills, where the repo-aware area validation test assumes Recommended follow-up: Either add a minimal
|
…icts Both branches added different id=1 evals and supplementary test cases. Combined all unique evals from each branch with sequential IDs: - fix-changelog: 17 evals (was 10 HEAD + 7 serverless) - review-changelog: 15 evals (was 9 HEAD + 6 serverless) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The frontmatter block was never closed, causing the CI validator to report "missing YAML frontmatter" for this skill. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Skill Eval Results
docs-fix-changelog
Overall: 78/79 expectations passed (99%) — 1 skipped, 1 fixture mismatch Skill instructions are thorough and the new feature-prefix evals (15–17) are well-supported by the explicit pattern in Step 4 item 1. The lone failure (eval 12) is a fixture mismatch — the prompt names docs-review-changelog
Overall: 71/71 expectations passed (100%) — 1 skipped Skill instructions cleanly cover all new feature-prefix eval cases (evals 13–15). The pattern distinction between true feature prefixes ("File upload:", "Dashboard:") and technical/API colons ("Authorization: Bearer") is explicitly modelled in Step 4. SummaryBoth skills pass at ≥99% when excluding the skipped area-validation eval (eval 11 in each skill). The skipped eval is a test environment limitation — the skill correctly specifies fallback behaviour when Recommended follow-ups:
|
This PR builds on #70 and addresses CI advice to add evals.json for the "review-changelog" and "fix-changelog" skills.