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.
An exceptionally actionable, well-sequenced review procedure: concrete commands, explicit gates, and feedback loops cover nearly every edge case, and supplemental detail is properly split into one-level-deep reference files. The only real costs are token weight from repeated policy statements and very dense paragraphs, and a body that carries more operational detail inline than a minimal overview would.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, domain-specific policy with essentially no explanation of concepts Claude already knows — every rule is procedural and non-obvious. It falls short of anchor 5 because several policies are stated more than once (credential stripping, verdict anti-hedging, the gh --jq shell note) and a few paragraphs (notably the Phase 3 test-check bullet) pack many rules into run-on form that could be tightened without losing content. | 4 / 5 |
Actionability | Fully executable throughout: exact gh commands with --jq filters, a complete paginated GraphQL query, credential-stripping command prefixes (env -u GH_TOKEN -u GITHUB_TOKEN pnpm --filter <pkg> test), exact file paths (.artifacts/factory-review/pr-<number>-<headSha>.md), exact verdict line formats, pre-publication gate commands (head -n1 <file>), and specific tool call sequences. Copy-paste ready with concrete failure modes and fallbacks. | 5 / 5 |
Workflow Clarity | Six explicitly sequenced phases with validation checkpoints and feedback loops at every risky point: pre-diff design discipline (Phase 1), bot-pending polling with a 60s/10min loop (Phase 2), environment-readiness checks and test-plan publication before execution (Phase 3), adversarial check and six named approval gates before any approve (Phase 5), and pre-publication file gates, head-moved checks, and address-then-retry on verdict-call rejection (Phase 6). This matches anchor 5's clear sequence with explicit validation and error-recovery loops. | 5 / 5 |
Progressive Disclosure | The bundle structure is well organized: the body is the core procedure, references/categories/README.md is a clearly signaled one-level-deep index of 16 category pages (verified present), and references/archaeology.md holds command recipes — both paths named explicitly in the body. It stops short of anchor 5 because the body itself is a 216-line procedure manual rather than an overview, carrying detailed operational minutiae (e.g. the Phase 6 routing/cross-check and author-identity edge cases) that could arguably live in reference files. | 4 / 5 |
Total | 18 / 20 Passed |