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.
The body is a strong operational document: a clearly sequenced, validation-heavy workflow with rollback, concrete CLI commands, and precise refusal criteria for a privileged mutation gate. Its weaknesses are token efficiency — rules are stated twice and design rationale is inlined — and a few executable details that are delegated to external shared-reference documents rather than shown inline.
Suggestions
Consolidate 'The non-negotiable rules' and 'Key Rules' into a single authoritative list (the latter currently restates the former nearly verbatim), keeping only cross-references.
Trim design-history narrative ('The earlier design trusted a jury_verdict: pass field...') to one line of rationale; the full history belongs in a project doc, not the skill body.
Inline a one-line example of the codex jury MCP call (JSON payload with model/config/sandbox/cwd) so Step 1 is copy-paste executable without reading reviewer-routing.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational (exact CLI flags, jury parameters, log schemas) but includes unnecessary material: the three non-negotiable rules are restated nearly verbatim in 'Key Rules', and sections like 'The acquittal is generated HERE, not relayed' and 'Provenance is a receipt, not an acquittal of correctness' carry design-history rationale that could be tightened. This matches anchor 3 (mostly efficient, some unnecessary explanation) rather than 4, where over-explanation would be only minor. | 3 / 5 |
Actionability | Guidance is largely executable: concrete commands (`python3 "$PROVENANCE" check --author ... --reviewer ...`, `stamp ... --trace-mode full`), exact jury call parameters (`model: gpt-6-astra`, `config: {"model_reasoning_effort": "ultra"}`, `sandbox: read-only`), and exact JSONL log shapes. It falls short of anchor 5 because the $PROVENANCE resolution chain and the codex MCP call payload are delegated to external documents/prose rather than shown as a copy-paste-ready example. | 4 / 5 |
Workflow Clarity | The Step 0→3 sequence is explicit with validation checkpoints throughout: refuse-default ('any failure ⇒ skip & report, never silently apply'), verify-before-mutation ('Missing or mismatched artifacts mean skip, even if the helper exited successfully'), a backup-before-edit rule, and a restore-on-failure feedback loop ('If apply, stamp, or logging fails, restore the target ... Do not claim LANDED for a partial transaction'). This matches anchor 5 (explicit validation, error-recovery loops) — the strongest dimension of the skill. | 5 / 5 |
Progressive Disclosure | The single-file body is well-sectioned and every external link (reviewer-independence.md, review-scope-limits.md, integration-contract.md §2, reviewer-routing.md, review-tracing.md, skill-governance.md) is one level deep and signaled with its purpose ('capability fallback is limited to explicit pre-thread model/effort errors'). It is 4 rather than 5 because no bundle files ship with the skill, key operational detail (helper resolution) lives entirely in an external contract, and the inlined duplicated rules could be consolidated or moved out. | 4 / 5 |
Total | 16 / 20 Passed |