Content
70%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 thorough, highly actionable paper-improvement loop with exemplary validation checkpoints and feedback loops. Its weaknesses are token efficiency (duplicated prompts and inline reference-grade material) and progressive disclosure (a monolithic file that should split out the edit-whitelist documentation).
Suggestions
De-duplicate the Round 1 and Round 2 review prompts (Steps 2 and 5) into a single shared template, parameterizing only the 'fresh zero-context review' instruction — this removes ~30 lines of near-verbatim repetition.
Move the ~100-line Edit Whitelist section (schema, resolution rules, glob semantics, forbidden-operation detectors) into a references/EDIT_WHITELIST.md and keep a brief summary plus link in SKILL.md to improve progressive disclosure.
Fix the duplicate 'Step 9' headers (both 'Document Results' and 'Summary' are numbered Step 9) and condense the repeated 'Empirical motivation' anecdotes into one short rationale per feature.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is operationally dense and mostly useful, but the ~30-line review prompt is duplicated verbatim between Step 2 and Step 5, the ~100-line Edit Whitelist section is inline rather than referenced, and repeated 'Empirical motivation' anecdotes pad the file — fitting 'mostly efficient but includes some unnecessary explanation or could be tightened'. | 3 / 5 |
Actionability | Concrete executable bash (latexmk, pdfinfo, grep), a runnable Python normalization heredoc, YAML schemas, and a regex-detector table give mostly copy-paste-ready guidance; the residual [VENUE] / [list all section .tex files] placeholders in the MCP prompts are minor gaps that keep it just below fully-executable anchor 5. | 4 / 5 |
Workflow Clarity | Steps 0–9 are explicitly sequenced with hard validation checkpoints — recompile verification ('0 undefined references, 0 undefined citations'), the Step 4.5 Restatement Regression Test, and the Step 8 location-aware format check with stop criteria and auto-fix patterns — plus feedback loops (validate → fix → re-validate), matching the anchor-5 example. | 5 / 5 |
Progressive Disclosure | Section structure is clear (## Context, ## Constants, ## Workflow), but the 581-line file is monolithic with no bundle files: the ~100-line Edit Whitelist docs and the duplicated review-prompt templates are content that clearly belongs in separate reference files, and the only external pointer (../shared-references/review-tracing.md) is not part of the bundle — fitting the 'content that should be separate is inline' anchor 3. | 3 / 5 |
Total | 15 / 20 Passed |