Content
52%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 defines a genuinely rigorous workflow — numbered scoping steps, exact git and script commands, explicit output templates, and enumerated fallback paths for every failure mode — and posting to PRs is properly guarded by an ask-first path. But at ~6,500 words it is heavily padded with 250–470-word single paragraphs of coordinator-only trace bookkeeping, and none of the 16 files it references exist in the bundle, leaving the lens protocol and reference structure unexecutable as shipped.
Suggestions
Ship the 16 referenced files (the six lens-*.md files, LENS-RETURN.md, NO-LENS.md, CHECK-TABLE.md, ACCEPTANCE.md, and the six scripts/) or remove the references — as bundled, every link and script invocation is broken, which is the main driver of the low progressive-disclosure score.
Move the trace/spawn bookkeeping (§3's tier-resolution paragraph, §6's relay emit vocabulary) into a reference file the coordinator reads on demand; it is not needed by the lens agents and inflates the always-loaded body past 6,000 words.
Break the 250–470-word single paragraphs into numbered step lists — each currently interleaves 5–10 concerns (tier resolution, Trace-Run lines, spawn.end outcomes, re-resolve fallbacks, raise timing) that a list would separate, improving both conciseness and workflow clarity.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~6,500 words the body is heavily padded: §3's tier-resolution paragraph alone runs ~470 words in a single block, three other paragraphs run 250–330 words each, and the 'axes are never merged' invariant is restated five times. It never explains concepts Claude already knows (ruling out anchor 1), but the verbosity and duplicated restatements clearly exceed anchor 2's 'noticeably verbose' only in degree, not kind — it sits at the severe end of anchor 2. | 2 / 5 |
Actionability | Concrete, literal commands appear throughout — 'git merge-base origin/<default-branch> HEAD', 'git log --oneline <merge-base>..HEAD', 'sh scripts/trace.sh begin review-pr subject=pr:#<N> || :', 'POLICY_BASE=<base> sh scripts/lens-slice.sh <base> <dir outside the tree>' — plus exact output templates and explicit decision rules. The gap keeping it from 5: the six referenced scripts are not shipped in the bundle, so those invocations cannot run as packaged. | 4 / 5 |
Workflow Clarity | The sequence is explicit (§0 steps 1–7 scope lock through §7 close) with real feedback loops: slicer exit-2/exit-3 fallbacks, spawn outcome=refused/unreachable/fail handling with a re-resolve recipe, verdicts recorded before posting, and the run closed only after raises. Anchor 5 is missed because 300–470-word single paragraphs interleaving many concerns make the otherwise sound sequence hard to follow. | 4 / 5 |
Progressive Disclosure | The design intends good disclosure — lens instructions in per-lens files, one-level-deep clearly-signaled links — but the bundle ships only SKILL.md: all 10 linked markdown files (lens-*.md, LENS-RETURN.md, NO-LENS.md, CHECK-TABLE.md, ACCEPTANCE.md) and all 6 scripts are absent, so every reference is broken, while coordinator-only trace bookkeeping that belongs in a reference file is inlined. | 2 / 5 |
Total | 12 / 20 Passed |