Content
62%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.
The body delivers an exceptionally rigorous, validated workflow with concrete commands and decision procedures, but at a heavy reading cost: monolithic paragraphs, repeated restatements, and inlined infrastructure code. The referenced companion files being absent from the bundle also leaves the progressive-disclosure structure only partly verifiable.
Suggestions
Split the trust-boundary section's ~70 lines of shell functions (vocab_checker, span_ok, prescreen_ok, checked_prescreen) into a scripts/ or references/ file, keeping only the call contract and the pre-screen end-to-end sequence inline.
Tighten the longest passages — rules 15–16 and steps 3–5 — into short normative statements with examples, and state the scratch-home cleanup once (in step 5) instead of three times.
Ship FACTORY.md, RISK.md, DOMAIN-RUBRIC.md, and RETRO-CANDIDATES.md alongside SKILL.md in the bundle so the four relative links resolve and the deferred content is actually available.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is noticeably verbose: rules 15 and 16 and steps 3–5 are 300–600-word single paragraphs with deeply nested em-dash clauses; the scratch-directory cleanup is stated at least three times (trust-boundary section, 'What the verdict means', step 5); and lines 154–156 re-explain in prose the exact logic the shell functions and their comments already specify. It avoids anchor 1 because nothing explains concepts Claude already knows — every passage is skill-specific — but well beyond anchor 3's 'some unnecessary explanation'. | 2 / 5 |
Actionability | Guidance is concrete and mostly executable: exact commands ('gh issue view "$PRD" --json body --jq .body', the mktemp templates, the full trace.sh emit call with every key), copy-paste shell functions (prescreen_ok, span_ok, checked_prescreen), exact stamp formats, and a first-hit-wins tier rubric. Anchor 5 is missed only because the reader-subagent spawn is deliberately left to the adapter ('the adapter names the command') and several oracle scripts live outside the bundle, leaving minor gaps the operator must fill. | 4 / 5 |
Workflow Clarity | The five-step procedure has explicit validation checkpoints and feedback loops throughout: checked_prescreen gates every read of the untrusted return, refused stamps are 'repaired and checked again, never shown', the coverage check enumerates exit codes 0–3 with re-runs after every change to a Covers: line, a mandatory human quiz gates publication, and cleanup paths exist for both completion and every stop condition. This matches anchor 5's sequence plus feedback loops plus error recovery. | 5 / 5 |
Progressive Disclosure | References are well signaled and one level deep — 'open [FACTORY.md](FACTORY.md) and run its check', 'the rubric is [RISK.md](RISK.md)', '[DOMAIN-RUBRIC.md](DOMAIN-RUBRIC.md)', '[RETRO-CANDIDATES.md](RETRO-CANDIDATES.md)' — and each is loaded exactly when needed. However, none of these four files is present in the evaluated bundle, and roughly 70 lines of trust-boundary shell functions that would naturally live in a scripts/ file are inlined in SKILL.md, which keeps it below anchor 5 and sits between anchors 3 and 4. | 4 / 5 |
Total | 15 / 20 Passed |