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 dense, highly executable operational runbook: every step carries exact commands, thresholds, and recovery paths, and the batch workflow has real validation checkpoints. Its weaknesses are at the margins — length with some deliberate redundancy, inlined sections that could live in reference files, one orphaned bundle file (babysit-round.md), and one ambiguously prefixed reference path (replies.md).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | There is no explanation of concepts Claude already knows — every line is an operational rule, command, or boundary — so it is well above the 'unnecessary explanation' anchor at 3. But at 686 lines it is not 'every token earns its place': critical rules are deliberately repeated ("Never wait for a fix branch, or end your turn to wait for one" appears in the Never list, step 4, and step 6) and a few multi-clause sentences (the `.ownership` rule in "Duplicate and competing PRs") could be tightened, matching anchor 4. | 4 / 5 |
Actionability | Fully executable throughout: a tools table with exact commands and flags (`factory-state report upsert "<key>" --json '<fields>'`), copy-paste-ready bash with realistic JSON payloads, prompt templates for fix and refactor branches, numeric gates ("a join under 70 doesn't attach"), exit-code handling ("exit 3 at the cap"), and a 10-minute timeout rule for branch starts. Common cases for every item kind are covered with specific handling. | 5 / 5 |
Workflow Clarity | Clear numbered sequence (steps 1–6: read the message, skip handled, triage and fold, start branches, record and close the loop, finish) with explicit validation and recovery: the finish checklist ("every claimed report has a digest card and every 👀 you added in live mode has its reply"), error branches for failed workers, launch-timeout recovery ("run `factory-branch list` and use the branch with this name"), and pre-action checks for the batch/destructive operations (duplicate check via `factory-state active` before any branch; "close nothing without both" before closing a PR). The batch-operation cap does not apply because validation checkpoints are present. | 5 / 5 |
Progressive Disclosure | References are one level deep and mostly well-signaled at point of use ("Read [confidence.md](references/confidence.md) before the first triage", fix-branch.md linked at each use, pr-review.md for review workers). But measured against the actual bundle: references/babysit-round.md is never reachable from the body, `references/replies.md` is cited without its `community-prs/` prefix and so dangles in this bundle, and long sections (Pull requests, Refactor findings) are inlined in SKILL.md rather than split — good structure, but more than minor gaps keeps it below the anchor-5 'overview with appropriately split content'. | 4 / 5 |
Total | 18 / 20 Passed |