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 highly actionable, well-sequenced workflow with exceptional validation and recovery machinery — format-check stop criteria, regression tests, and bias-guard rules are best-in-class. Its weaknesses are duplication (the reviewer prompt repeated nearly verbatim), the monolithic 634-line body that inlines spec content better suited to reference files, and placeholder-riddled templates that stop short of copy-paste readiness.
Suggestions
Replace the near-verbatim duplicate reviewer prompts in Steps 2 and 5 with a single 'Reviewer prompt template' section (the Round 2 variant's only addition is the fresh-context preamble), saving ~50 lines.
Move the ~105-line edit-whitelist specification (schema, resolution rules, glob semantics, detectors, behaviors) into a references/edit-whitelist.md file and keep only a short summary plus the link in SKILL.md.
Complete the Step 4.5 Python snippet so it actually compares normalized theorem blocks (e.g., extract \begin{theorem}...\end{theorem} spans per file and diff them) instead of ending at a comment, and fix the duplicated 'Step 9' section numbering.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The bulk is dense operational specifics, but the ~50-line reviewer prompt is duplicated nearly verbatim in Steps 2 and 5, the edit-whitelist gate is restated in Steps 3, 6, and Key Rules, and long empirical anecdotes pad several sections. Anti-drift: above anchor 2 because most sections carry unique actionable content rather than padded fluff. | 3 / 5 |
Actionability | Concrete bash commands (latexmk, pdfinfo, grep overfull classification), full reviewer prompt templates, regex detector tables, a YAML whitelist schema, and a Python normalization snippet. Minor gaps — [VENUE]/[list figure files] placeholders, the Python block ending in a comment instead of a complete comparison, and the pseudo-YAML spawn_agent invocation — keep it below fully copy-paste-ready anchor 5. | 4 / 5 |
Workflow Clarity | Steps 0–9 are explicitly sequenced with validation checkpoints throughout: verify 0 undefined references, a restatement regression test that blocks completion, location-aware format-check stop criteria, crash-recovery state persistence, and optional human checkpoints with response-parsing rules. Feedback loops (fix → revalidate → block) are exactly what the rubric's top anchor rewards; the duplicated 'Step 9' numbering is a cosmetic flaw only. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent; the only pointers go to ../shared-references/ outside this bundle). The 634-line body inlines large content blocks that clearly belong in separate files — the ~105-line edit-whitelist spec and the reviewer prompt template — despite otherwise good section structure. Anti-drift: above anchor 2 because structure is solid and clearly signaled, below anchor 4 because significant content should be split out. | 3 / 5 |
Total | 15 / 20 Passed |