Content
85%Weight 40%Scale 1-3Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A highly actionable, validation-gated orchestration skill with executable commands, a clear sequenced workflow, and a sound one-level-deep reference split. Its one weakness is conciseness: several load-bearing invariants are restated verbatim across sections and bash setup boilerplate is duplicated, so the 247-line body is longer than necessary without adding clarity.
Suggestions
State each core invariant once — e.g. the 'don't serialize comments and CI' principle and 'ending the loop on a needs-human residual is the primary failure mode' — in the core-principle section, then reference it from Step 2/3/5 instead of restating full paragraphs 3–4 times.
Extract the repeated bash setup boilerplate (SKILL_DIR/SCRATCH_ROOT/STATE_DIR/PY detection, ~7 lines duplicated across ~5 command blocks) into watch-loop.md or a single canonical snippet the later blocks assume, cutting a large fraction of the body length.
Move the detailed merge-ready wake-protocol branches and the full review-still-expected guard into references/watch-loop.md, keeping only the decision rule and the stop ceilings inline in SKILL.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is accurate and operationally dense with no basic-concept padding, but core invariants are restated multiple times — the 'don't serialize comments and CI' principle appears in the intro, the core-principle block, and Step 2, and 'ending the loop the moment one item needs a human is the primary failure mode' is repeated ~4 times — so it 'could be tightened' rather than being lean with every token earning its place. | 2 / 3 |
Actionability | Provides fully executable bash with exact flag names ('pr-snapshot snapshot --pr <N> --repo ... --start-invocation --invocation-budget-seconds <seconds>'), exact gh commands ('gh run rerun <run-id> --failed -R <host>/<owner>/<repo>'), and an exact REST endpoint ('PUT /repos/{owner}/{repo}/pulls/{number}/update-branch' with expected_head_sha) — copy-paste ready, matching the score-3 anchor. | 3 / 3 |
Workflow Clarity | Clear sequence (Steps 1–5 plus an 8-step ordering invariant) with explicit validation checkpoints — head-SHA capture and stale-SHA cancellation, OID revalidation before mutation, the managed-stack atomicity gate, abort-on-conflict residuals — and feedback loops (fix → re-validate), matching the score-3 anchor; the rubric's destructive/batch cap does not apply since validation is pervasive. | 3 / 3 |
Progressive Disclosure | Clean one-level-deep reference structure: watch-loop.md is a genuine terminal reference (no nested refs) signaled at clear points ('Read references/watch-loop.md for the mechanics / state schema / edge cases'), and scripts/pr-snapshot is the referenced executable helper — matching the score-3 anchor of well-signaled one-level-deep references with easy navigation; the inline length is a conciseness concern, not a disclosure-structure failure. | 3 / 3 |
Total | 11 / 12 Passed |