Content
92%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.
An exemplary reference skill: signal/cause tables, an executable worked example, a decision-tree checklist with an explicit re-measure validation loop, and a clean split of deep code fixes into three verified one-level-deep reference files. The only improvement is trimming the terminology-provenance note and the test-size correlation aside, which add background rather than actionable content.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with actionable signal — per-pattern signal/cause tables, named APIs and flags, a quick-reference table, and a worked example with before/after code — so most tokens earn their place. It is not 5 because a few passages explain background Claude does not need: the terminology-provenance note ("'flaky test' is a practitioner-emergent term popularized by the Google Testing Blog... ISTQB does not maintain a canonical entry") and the "~4.2M tests" test-size correlation aside; it is not 3 because these are minor trims, not padded sections. | 4 / 5 |
Actionability | Concrete, executable guidance throughout: exact commands and flags (`jest --randomize`, `pytest --random-order`, `TZ=UTC`, `--testTimeout`), specific APIs (`page.clock.install()`, `sinon.useFakeTimers()`, `faker.seed(N)`, `testInfo.workerIndex`, `page.route()`), and a copy-paste-ready before/after TypeScript diff in the worked example. The quick-reference table covers all eight patterns with a key fix and primary API, matching the 5 anchor's requirement that specific examples cover the common cases. | 5 / 5 |
Workflow Clarity | The triage decision tree is an explicit sequenced checklist, and the 'Code-level fixes' section defines the workflow — classify (inspection or `flake-axis-bisection`), apply the smallest change, re-measure, quarantine if blocking — with an explicit validation checkpoint ("re-run at an N chosen from the failure rate you are willing to ship... a clean 0/20 does not prove the flake is gone") and a feedback loop. The worked example walks all three steps including the N=300 re-measure, satisfying the 5 anchor (explicit validation, feedback loops, checklist) rather than the 4 anchor where checkpoints have minor gaps. | 5 / 5 |
Progressive Disclosure | The split is well-judged: pattern signals and brief remediations stay in SKILL.md while the per-pattern code fixes live in three reference files, each linked with a clear pattern-range description ("Patterns 1-4", "Patterns 5-6", "Patterns 7-8") under 'Code-level fixes'. Verified against the actual bundle: all three referenced files exist in references/, and each is one level deep (external doc citations only, no nested .md references), matching the 5 anchor for well-signaled one-level-deep references with content appropriately split. | 5 / 5 |
Total | 19 / 20 Passed |