Content
77%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A highly actionable, well-sequenced auditor prompt with concrete detection patterns, fix examples, and verification steps. Its main weakness is conciseness and progressive disclosure: it is a long single-file wall of text with some explanatory padding that could be trimmed or offloaded to reference files.
Suggestions
Trim explanatory 'Issue/Why it matters' prose and the OS27 rationale to one line each; Claude already understands flaky-test and concurrency trade-offs, so lead with the grep pattern and fix.
Extract the per-category grep patterns and fix code into a reference file (e.g. references/anti-patterns.md) and keep SKILL.md as a phased overview with one-level-deep links, improving progressive disclosure and token budget.
Consolidate the Phase 3/4 question and compound-finding tables into a single reference table to reduce repetition between the two phases.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient and actionable, but at ~340 lines it carries explanatory prose Claude doesn't need (e.g. 'Issue: Arbitrary waits cause timing-dependent failures, especially in CI' and the long OS27 rationale paragraph) that could be tightened without losing clarity. | 2 / 3 |
Actionability | Provides concrete grep regexes, exact file globs, executable Swift fix examples, and precise detection/flagging criteria — copy-paste ready guidance throughout, matching the top anchor. | 3 / 3 |
Workflow Clarity | Sequenced across Phase 1–5 with explicit verification ('read surrounding context to verify it's a real issue before reporting'), cross-reference tables, a false-positive checklist, and a defined output format — clear sequence with validation checkpoints for a batch audit. | 3 / 3 |
Progressive Disclosure | Sections are well-organized, but no bundle files exist and the entire auditor lives inline in one ~340-line SKILL.md; external skill references (axiom-testing, axiom-ai, test-failure-analyzer) are signaled but inline content that could be split keeps it below the top anchor. | 2 / 3 |
Total | 10 / 12 Passed |