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 concrete orchestration skill with exceptional validation and feedback-loop coverage (regression tests, blocking criteria, state recovery, audit logging). Its weaknesses are the inlined-then-duplicated review prompt and whitelist spec, which inflate token cost and would benefit from reference-file extraction, plus one incomplete inline script.
Suggestions
Replace the duplicated Round 2 review prompt in Step 5 with a reference to the Step 2 prompt template plus only the Round 2 deltas (fresh-thread instruction, zero-context line, 'Ignore any prior review rounds' paragraph), saving ~50 lines.
Move the edit-whitelist spec (schema, resolution rules, glob semantics, detector table, logging schema, ~100 lines) into a reference file such as references/edit-whitelist.md and keep a short summary, the detector table, and a link in SKILL.md.
Complete the Step 4.5 inline Python beyond the normalize() helper — provide the theorem-block extraction and comparison logic (or explicitly route to /proof-checker --restatement-check as the executable path) so the regression gate is runnable as written.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | All content is task-specific with no basic-concept padding, but the ~50-line reviewer prompt is duplicated verbatim in Steps 2 and 5, and multiple 'Empirical motivation'/'Rationale' blocks could be tightened or merged. Fits anchor 3 (mostly efficient, could be tightened) rather than 2 because the verbosity is domain-relevant, not generic filler, and not 4 because the prompt duplication is a large avoidable block. | 3 / 5 |
Actionability | Nearly every step is copy-paste executable (latexmk commands, pdfinfo/grep format checks, YAML whitelist schema, regex detectors, MCP call configs), but the Step 4.5 inline Python defines only normalize() and leaves the actual theorem-block comparison as a comment, and [VENUE]/$PAPER_DIR placeholders require substitution. Anchor 4 (mostly executable, minor gaps) rather than 5 because the regression-test script is incomplete. | 4 / 5 |
Workflow Clarity | Steps 0-9 are explicitly sequenced with validation checkpoints throughout: 'Verify: 0 undefined references, 0 undefined citations', the Step 4.5 restatement regression test, location-aware format-check stop criteria, whitelist rejection logging, JSON state recovery for compaction, and HUMAN_CHECKPOINT feedback branches. Matches the anchor-5 example's validate->fix->retry rigor; the duplicate 'Step 9' heading numbering is cosmetic. | 5 / 5 |
Progressive Disclosure | The bundle contains zero reference/script files; the entire 710-line implementation lives in SKILL.md, including an ~100-line edit-whitelist spec and the full review prompt template that clearly belong in separate files. External references (shared-references/*, tools/extract_paper_style.py, skills/kill-argument/) are clearly signaled but outside the bundle; anchor 3 (content that should be separate is inline) rather than 2 because sections are well-organized and references are not buried. | 3 / 5 |
Total | 15 / 20 Passed |