Content
82%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tightly written, opinionated body that assumes Claude's competence and consistently teaches judgment rather than basics. The gaps are mechanical rather than conceptual: no executable example to anchor the test shape, and all guidance inlined in one file rather than split across references.
Suggestions
Include one minimal executable example (e.g., a short pytest snippet of a session-start, one turn, and a structural assertion) marked as 'check current signatures via reading-livekit-docs' — it would anchor the described shape in concrete code without going stale.
Consolidate the placement framing: the intro and the 'Where this sits' section both argue that these tests are cheap, fast, and run every commit; stating it once would trim a few lines.
Consider moving the 'Three kinds of evidence, kept distinct' deep-dive into a references/ file, keeping SKILL.md as a leaner overview that points to it — this would complete the progressive-disclosure structure the other sections already gesture toward.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body never explains concepts Claude already knows (no 'what LiveKit is', no pytest tutorial) and every section teaches non-obvious judgment: 'Mock failures as well as successes', 'Mocking only changes execution. The model still sees the real tool schemas', the three-kinds-of-evidence taxonomy. The only trimmable material is slight framing duplication between the intro and 'Where this sits', which stays on the 5 anchor ('every token earns its place') rather than dropping to 4. | 5 / 5 |
Actionability | The guidance is concrete and specific — the test shape ('start a test session with the agent under test, run one user turn, and assert on the events that turn produced'), a worked intent example ('tells the user the booking is confirmed and gives the time'), and a prioritized what-to-test list — but it contains no executable code or commands at all; all API mechanics are deferred to 'reading-livekit-docs' and 'the testing docs page'. That deferral is justified (helper names change), and the rubric's instruction-only note says not to penalize absent code when guidance is actionable, so it sits above the 3 anchor (guidance is complete, not pseudocode) but below 5 (nothing is copy-paste ready). | 4 / 5 |
Workflow Clarity | A clear sequence is given: look up the API surface first ('look them up with reading-livekit-docs before writing'), start a session, run one turn, 'walking that sequence in order, asserting on each event', then 'Close the turn.' The anti-pattern sections provide error guidance ('If the agent can only pass by misbehaving, fix the test'), but there is no explicit validate/fix/retry feedback loop for a failing test run, which keeps it below the 5 anchor while clearly above 3. | 4 / 5 |
Progressive Disclosure | The file is well-sectioned with clear headers, an explicit pointer for detail ('use the testing docs page for every API detail this skill leaves out'), and a closing 'Related skills' section — all references are clearly signaled and one level deep. However, all ~125 lines of guidance live inline in a single file with no bundled reference split (no references/ or scripts/ exist), and the under-50-lines simple-skill exception does not apply, so it falls just short of the 5 anchor's 'content appropriately split'. | 4 / 5 |
Total | 17 / 20 Passed |