Content
85%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.
A well-engineered skill body: precise step sequencing, executable commands, real feedback loops, and exemplary use of one-level-deep reference files. The only notable weakness is redundancy — the same hard rules are restated in the philosophy, contract, and hard-nevers sections.
Suggestions
Consolidate the triple statement of the memory rule (E7, the "Invocation & output contract" bullet, and the first "Hard never"): state it once as a Hard never and have E7 and the contract reference it in a single line.
Similarly fold the repeated "never hand work back to the loop" rule (E9, Step 7 closing, and the Hard never) into a single authoritative location, keeping only one-line cross-references elsewhere.
The E7 paragraph's deep dive into /learn's P7 semantics belongs in references/verdict.md (where the merge → /learn path is already explained); replacing it with a one-line pointer would cut ~10 lines without losing the rule.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body carries dense, genuinely novel workflow knowledge with no padding about concepts Claude already knows, but the same invariants are restated two to three times: E7's memory rules appear in the philosophy section, the "Invocation & output contract", and "Hard nevers"; E3, E8, and E9 are similarly repeated. This fits anchor 3 ("mostly efficient but... could be tightened") — more than minor trimmable redundancy, but not the majority-padded anchor 2. | 3 / 5 |
Actionability | Fully executable guidance throughout: exact commands ("git checkout --detach origin/feature/<f>", "git push origin HEAD:feature/<f>", "gh pr view <pr> --json reviews,comments", "git checkout main"), an ordered reading list of concrete paths (prds/<f>/prd.md, run-prd-test.sh, specs/<f>/mainspec.md), and copy-paste verdict commands in references/verdict.md. As an instruction-only skill this matches the anchor-5 standard. | 5 / 5 |
Workflow Clarity | Steps 0–8 are clearly sequenced with explicit validation checkpoints (clean-tree preflight, the Step 6 understanding gate, human go-ahead required before merge/close) and a genuine feedback loop in the fix-then-merge path (push → reviewer re-runs → wait for HARNESS_REVIEW_CLEAN → merge). The "Hard nevers" section functions as a checklist. This matches anchor 5. | 5 / 5 |
Progressive Disclosure | The body is a clear overview that moves bulk detail into two well-signaled, one-level-deep references (references/walkthrough.md and references/verdict.md — both present in the bundle and delivering what their intros promise), each introduced with its scope and hackable seams, with no nested reference chains or orphan pointers. Matches anchor 5. | 5 / 5 |
Total | 18 / 20 Passed |