Content
88%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 highly operational, fail-closed scheduler manual: executable commands with flag-level variants, a mandated pre-spawn queue table, stop-on-failure validation gates, and a well-signaled one-level reference to the fan-out pool. The only real costs are systematic repetition of the write-prohibition rules and dense inline detail in the scheduled-sweep section that could be offloaded to a reference file.
Suggestions
Consolidate the write-prohibition rules (stated in the intro, Activation card rules, Invocation, Selection, and Hard nos) into one canonical section and reference it, reclaiming redundant tokens.
Move the scheduled-sweep eligibility details (exclusion labels, evidence requirements, five-item cap semantics) into a short reference file and keep only the invocation contract inline.
Add one filled example row to the keep-set/drop-set table spec so the mandatory format is unambiguous on first use, and fix or contextualize the ../autopilot-issue-triage-worker/SKILL.md relative link.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, imperative, and free of concept explanations Claude already knows, but the write-prohibition ("never comments, labels, assigns, or requests reviewers") is restated across ~5 sections and the Invocation "Ownership writes" list describes writes ("Code (accepted implementation): assign the implementing user") this scheduler never performs; this is more than zero padding but stays minor, fitting anchor 4 rather than anchor 5. | 4 / 5 |
Actionability | Fully executable guidance throughout: copy-paste commands ("python <this-skill>/scripts/fetch_queue.py --kind issue --mode sweep --repo microsoft/apm > batch.json"), exact flag variants (--mode dispatch --number N, --records-json, --labels-json, --scheduled-sweep), field-by-field activation card templates, and an exact queue-table column spec, covering the default, named-issue, and scheduled-sweep cases. | 5 / 5 |
Workflow Clarity | The 5-step Procedure is clearly sequenced and saturated with explicit validation gates for this batch fan-out: "Missing -> stop" (probe), "Missing table, missing column, or blank rationale -> stop. Do not spawn.", "A helper failure stops the run", "Incomplete evidence fails the sweep", plus the Exit receipt and final report; validation is present so the batch-operation cap does not apply, matching anchor 5. | 5 / 5 |
Progressive Disclosure | Against the actual bundle, assets/fan-out-pool.md is explicitly linked ("Load [assets/fan-out-pool.md](assets/fan-out-pool.md)") one level deep and scripts/fetch_queue.py, scripts/triage_state.py, and assets/label-contract.json are named as canonical owners; however the dense scheduled-sweep eligibility rules are inline rather than split to a reference, the queue-table spec lacks a filled example row, and the relative link ../autopilot-issue-triage-worker/SKILL.md has no target in this bundle, so anchor 4 fits rather than 5. | 4 / 5 |
Total | 18 / 20 Passed |