Content
45%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.
The body defines a genuinely well-sequenced iterative workflow with strong validation gates, but it is bloated by heavy repetition of the same style rules, monolithic inlining of material that belongs in reference files, and scripts with placeholders and gaps that never materialize several promised output artifacts. It reads as an over-stuffed prompt rather than a lean, executable skill.
Suggestions
State the color palette, arrow standards, and 科研风格 rules exactly once — move the full style guide into references/style-guide.md and have the Step 1 template and Step 3/5 checks reference it instead of re-inlining it.
Close the execution gaps: make Step 4 write review_log.json and copy the accepted render to figure_final.png, generate latex_include.tex in Step 8, and drive the iteration counter from a variable instead of the hardcoded `iteration = 1`.
Move the Step 2/3/4 bash scripts into a scripts/ directory (parameterized by prompt file and output path) so the SKILL.md body shrinks to the workflow, rules, and review template.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~690-line body restates the same style guidance four times (style guide section, Step 1 prompt template, Step 3 verification request, Step 5 checklist) — the identical color palette and arrow rules appear repeatedly — plus heavy ❌/✅/═══ decoration and bilingual redundancy. It avoids explaining basic concepts Claude already knows, so it sits above the severely-verbose anchor but is clearly noticeably verbose. | 2 / 5 |
Actionability | Concrete curl/python scripts exist, but they contain placeholder holes ("INITIAL_PROMPT='[Codex fills in the detailed prompt here]'"), a hardcoded "iteration = 1 # Codex increments this each iteration", fragile '''$VAR''' shell-into-Python quoting that breaks on quotes in the prompt, and no step that actually produces the declared figure_final.png, latex_include.tex, or review_log.json. This matches "some concrete guidance but incomplete; missing key details" rather than the mostly-executable anchor above. | 3 / 5 |
Workflow Clarity | The Step 0–8 sequence is clear with an explicit decision point, a genuine feedback loop (score < 9 → targeted improvement prompt → back to Step 2), a MAX_ITERATIONS guard, and a detailed review checklist — near the top anchor. It falls short of 5 because the declared output artifacts (figure_final.png copy, review_log.json, latex_include.tex) are never produced by any scripted step, a real coherence gap. | 4 / 5 |
Progressive Disclosure | The skill is a monolithic single file with no references/, scripts/, or assets/ bundle; the 200+ line CVPR style guide and the long prompt templates clearly belong in separate reference files. Section headers provide some structure, placing it above the unstructured anchor but squarely at "content that clearly belongs in separate files is inlined". | 2 / 5 |
Total | 11 / 20 Passed |