Content
56%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 skill body is supremely actionable and rigorously gated, but it is buried in one of the most verbose skill documents possible — incident narratives, historical justifications, and edge-case essays inflate it to ~100K tokens, defeating the token budget its own design philosophy implies. The reference-file gating pattern that works for posting/persistence/aone should have been applied to most of the remaining body.
Suggestions
Move the per-incident rationale out of SKILL.md entirely: every "(measured; DESIGN.md — …)" narrative and dated war story (PR #9729, #8368, the 22-minute serial verification) already lives in DESIGN.md for maintainers — keep at most a one-line pointer, cutting a large fraction of the 416KB body.
Apply the existing posting/persistence/aone reference pattern to the remaining monolith: split Step 1's argument-parsing and prose-resolution rules, Step 4's witness/capability catalog (probe, base-tree, impact sweep, isolation-by-elimination), and Step 5's reverse-audit scheduling into gated one-level-deep reference files loaded only when the run's verdict needs them.
Convert multi-thousand-character single paragraphs into short imperative rules with a one-line rationale (or a bullet list); executable rules like the cleanup prefix-collision contract are currently buried mid-paragraph in narrative, which is what makes an otherwise clear 9-step pipeline hard to trace.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is ~416KB (~1234 lines averaging 336 characters, with single paragraphs over 7,000 characters) — a context-window-consuming monolith. Nearly every rule carries an inline incident narrative ("measured; DESIGN.md — The transcribed argument file", "measured on PR #9729") and multi-clause historical justification, i.e. heavily padded, and it explains at length concepts a competent model knows (heredoc termination, POSIX parameter expansion, why git diff misses untracked files). This matches the severely-verbose/heavily-padded anchor, not merely "several padded sections". | 1 / 5 |
Actionability | Guidance is fully executable throughout: exact copy-paste commands ("${QWEN_CODE_CLI:-qwen}" review parse-args --stdin --out .qwen/tmp/qwen-review-parse-args.json), exact file paths, exit-code semantics (0/3/4/5/6/7 with prescribed responses), explicit JSON verdict field meanings, and concrete output templates (the "Review complete: <target> — <disposition>" contract). This matches the copy-paste-ready top anchor. | 5 / 5 |
Workflow Clarity | The 9-step sequence is clearly ordered with deterministic gates, explicit validation (verifier shards, convergence pairs, exit-code-driven fail-closed handling) and error-recovery feedback loops — well above the missing-checkpoint anchors. It falls short of the top anchor because at least five orthogonal axes (effort level × topology × platform × review profile × target type) are interwoven through continuous 3,000–7,000-character prose with no scannable checklists or per-path summaries, making any single execution path genuinely hard to trace. | 4 / 5 |
Progressive Disclosure | The three existing references (references/posting.md, references/persistence.md, references/aone.md) are real files, one level deep, and exceptionally well-gated ("Load it when, and only when, posting is live for this run") — better signaling than the 3 anchor describes. But the overwhelming bulk of conditional detail (Step 1's ~330 lines of parsing/prose-resolution rules, Step 4's witness/capability catalog, Step 5's audit scheduling) is inlined in a monolithic body that dwarfs the reference files (416KB vs 111KB), so content that clearly belongs in separate files is inline. | 3 / 5 |
Total | 13 / 20 Passed |