Content
77%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A highly actionable, well-sequenced workflow with strong validation checkpoints and feedback loops, but it is verbose and monolithic. Tightening duplicated prompt templates and splitting large blocks into reference files would meaningfully improve it.
Suggestions
Deduplicate the Phase A prompt templates: define the shared senior-reviewer instructions once and have the hard/nightmare variants reference or prepend only their deltas (Reviewer Memory, debate steps) instead of repeating the full prompt.
Move the large per-difficulty prompt blocks, the Debate Protocol, and the Reviewer Memory template into separate reference files under references/ and link to them one level deep, reducing the SKILL.md body to an overview plus the core workflow.
Collapse the repeated 'Please act as a senior ML reviewer... Be brutally honest' framing and the round-2+ template's restated format into a single canonical prompt skeleton referenced by all phases.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is largely procedural and avoids explaining concepts Claude already knows, but it is long with duplicated prompt templates (the medium and hard Phase A prompts differ only by a prepended memory block, 'Be brutally honest' recurs, and the round-2+ template restates the format), so it could be tightened. | 2 / 3 |
Actionability | Provides fully executable guidance — concrete mcp__codex__codex calls with config, a copy-paste codex exec bash block, an exact REVIEW_STATE.json schema, and verbatim markdown templates — leaving little ambiguous. | 3 / 3 |
Workflow Clarity | Sequences a clear Init → Loop (Phase A–E) → Termination flow with an explicit STOP CONDITION (score >= 6 AND verdict ready/almost), state-persistence checkpoints after each round, and a review→fix→re-review feedback loop for error recovery. | 3 / 3 |
Progressive Disclosure | No bundle files exist (references/scripts/assets are absent) and the body is a single ~440-line monolith with well-organized sections but everything inline; the large per-mode prompt templates and debate protocol are candidates for one-level-deep reference files. | 2 / 3 |
Total | 10 / 12 Passed |