Content
73%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 tightly engineered, highly actionable process specification with an exemplary five-phase workflow and explicit validation gates throughout. The main weaknesses are redundancy (a metadata comment block and repeated rules that could be trimmed) and heavy reliance on roughly twenty reference files that are not present in the reviewed bundle, some of which hold definitions (e.g., HP/NP/BP/EP) the body uses without defining.
Suggestions
Trim the duplication: remove the CAPABILITIES_SUMMARY HTML comment block (its routing content is already in the body's Collaboration and Trigger Guidance sections) and state the 'Attest writes no code' rule once in Boundaries, citing it elsewhere instead of restating it.
Define the scenario type abbreviations HP/NP/BP/EP inline (a one-line parenthetical in the priority-minimums sentence) so the body's most-used convention is self-contained without requiring reference/bdd-generation.md.
Consolidate recurring content — the Operating Modes table and Recipes table overlap in mode-binding detail, and the fix-prompt suppression conditions appear in Core Contract, Output Requirements, and LLM Fix Prompt Generation; state each once and cross-reference.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and telegraphic with no explanations of concepts Claude already knows, but a meaningful fraction is trimmable duplication: the 30-line CAPABILITIES_SUMMARY HTML comment repeats routing info already in the body, 'Attest writes no code' and the BOUNDARIES.md citation recur three-plus times, and operating-mode and verdict details appear in multiple sections. It fits 'mostly efficient but could be tightened' better than anchor 4, where over-explanation would be only minor. | 3 / 5 |
Actionability | Concrete ID schemes (AC-{FEATURE}-{NNN}, SC-{criterion_id}-{type}-{NNN}, PRB-{category_code}-{NNN}), numeric gates (confidence >= 0.8 / 0.5-0.8 / < 0.5; 20+ criteria; >30% ambiguity; <= 3 HIGH PARTIAL; >= 90% traceability), verdict tables, a fixed report section order, and named fix-prompt verbs make the guidance mostly executable. Minor gaps remain: the HP/NP/BP/EP scenario types are used in the priority minimums but never defined in the body, and key mechanics (evidence schema, confidence computation) defer to reference files absent from the bundle. | 4 / 5 |
Workflow Clarity | The INGEST -> EXTRACT -> GENERATE -> VERIFY -> ATTEST pipeline is laid out in a table with per-phase goals, required outputs, and files to read, backed by explicit validation checkpoints: ingest confidence gates, 'Confidence < 0.5 -> NOT_TESTED, never PASS', Ask-First triggers with timing (SPEC_MISSING, SCOPE_SELECTION, AMBIGUITY_CRITICAL, REJECT_CRITICAL), mode auto-detect rules, and subcommand dispatch. As a read-only verification skill, no destructive/batch validation cap applies. | 5 / 5 |
Progressive Disclosure | The design is strong: a Reference Map with a 'Read this when' column, a per-phase Read column, inline thresholds with detail deferred, and one-level-deep signaling. However, none of the ~20 referenced files (reference/*.md, _common/*.md) exist in the reviewed bundle, so the split cannot be verified, and the sheer number of external dependencies burdens navigation — good structure with gaps rather than fully clean. | 4 / 5 |
Total | 16 / 20 Passed |