Content
77%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The content is a clear, actionable multi-round review workflow with concrete tool names, an executable install command, and explicit polling/convergence checkpoints. Its weaknesses are redundant prompt content that bloats the token budget and a monolithic structure with no progressive disclosure to reference files.
Suggestions
De-duplicate the initial-review prompt: keep the full template in the Prompt Templates section and reference it from Step 2 instead of restating it inline.
Tighten the Key Rules section by removing items already implied by the workflow steps (e.g., "Send comprehensive context in Round 1" duplicates Step 1).
Move the prompt templates into a referenced file (e.g., references/prompts.md) with a one-line pointer so SKILL.md stays a lean overview and improves progressive disclosure.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly efficient and avoids explaining concepts Claude already knows, but the "senior ML reviewer (NeurIPS/ICML level)" prompt appears in both Step 2 and the Prompt Templates section, and the Key Rules restate guidance already in the workflow, so it could be tightened. | 2 / 3 |
Actionability | It gives concrete MCP tool identifiers (review_start, review_reply_start, review_status), a copy-paste install command, a precise polling protocol (save jobId, poll with bounded waitSeconds until done=true, reuse threadId), and specific prompt templates — fully actionable guidance. | 3 / 3 |
Workflow Clarity | Five steps are clearly sequenced with an explicit async-completion checkpoint (poll review_status until done=true), convergence stop-conditions in Step 4, and an iterative dialogue feedback loop for responding to criticisms. | 3 / 3 |
Progressive Disclosure | The body is well-organized into sections but is a single ~100-line monolithic file with no reference files and inline prompt templates that could be split out, so it does not reach the one-level-deep reference structure of the level-3 anchor. | 2 / 3 |
Total | 10 / 12 Passed |