Content
77%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 exceptional actionability and workflow discipline — executable commands, exact artifact contracts, and validation gates at every step — and nearly all of its content is hard-won, non-obvious guidance rather than padding. Its weaknesses are size and placement: a ~2,200-line monolith with one reference file, containing repeated measured examples and whole operational recipes that belong in progressive-disclosure reference files.
Suggestions
Move the self-contained operational recipes — the Docker/model-proxy container recipe, Watch mode setup, and the maintainer-driven local publish rules — into references/*.md files, keeping SKILL.md to the entry paths, environment contract, scope selection, method summary, and artifact contract.
Deduplicate measured examples that appear in multiple sections (e.g. the parked-Turn-2 cancel example in both Scope selection item 1 and Wire-oracle harnesses' 'Parameterise scenarios'), keeping one canonical instance and cross-referencing it.
Compress the A/B methodology's ~30 heuristic bullets to a one-line trigger rule each, relocating the extended "Measured example" narratives to a methodology reference file loaded only when scoping A/B work.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~2,200 lines the body is dense with genuinely non-obvious, measured lessons and does not explain basics Claude already knows, but it is well past lean: measured examples repeat across sections (e.g. the parked-Turn-2 cancel example appears in both Scope selection and Wire-oracle harnesses, kept byte-identical by the file's own admission), and the publish/guard/watch policy prose could be substantially tightened. Anchor 3 ("mostly efficient but… could be tightened") fits better than 4, where trimming needs would be minor. | 3 / 5 |
Actionability | Guidance is fully executable end to end: copy-paste bash blocks for the container recipe and proxy, exact commands ("git worktree add tmp/base-tree <base>", "node scripts/verify-capture.mjs --out … -- <cmd>"), exact artifact contracts (verdict.txt "exactly one word", assertions.json field shapes with attribution enums), and concrete cell-table shapes to aim for ("5/9 flip from broken to fixed"). | 5 / 5 |
Workflow Clarity | The process is explicitly sequenced with validation checkpoints throughout: "Scope selection (do this before running anything)", the isolation gate with exit-code semantics ("treat exit 4 as 'this is not a boundary'"), the checker run before publishing, A/A controls for environmental failures, and "Fail loud" as a hard rule. Feedback loops (validate → fix → re-validate, re-measure never diff) are present at every risky stage. | 5 / 5 |
Progressive Disclosure | The single bundle reference (references/android.md) exists and is clearly signaled at the point of need ("read references/android.md before scoping"), and section structure is strong — but the remaining ~2,200 lines, including the full Docker/model-proxy recipe, watch-mode setup, local publish rules, and ~30 methodology heuristics with extended narratives, are inlined in SKILL.md where reference files would serve. Anchor 3 ("content that should be separate is inline") applies; structure is too good for anchor 2. | 3 / 5 |
Total | 16 / 20 Passed |