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.
The workflow engineering is excellent — sequenced phases, checkpoints, error loops, and executable commands dominate the body. Weaknesses are length and monolithic structure: large templates belong in reference files, the reviewer-routing preamble is confusing and points to a file not in the bundle, and some guidance duplicates across sections.
Suggestions
Split the full beamer template, talk-type slide tables, and the Phase 8 talk-script example into files under references/ and link them from SKILL.md, keeping the body as an overview (progressive_disclosure is the lowest-scoring dimension).
Rewrite or relocate the reviewer-routing blockquote: it is dense, references ../shared-references/reviewer-backends.md which is absent from this bundle, and duplicates REVIEWER_MODEL routing details that belong in a reference file.
Deduplicate the slide-count guidance (Constants table vs. Phase 1 tables) and compress the Phase 8 markdown example to a short skeleton to reduce token cost.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — tables, templates, and commands rather than concept explanations — but it could be tightened: the dense reviewer-routing blockquote at the top, slide-count guidance duplicated between Constants and Phase 1, and the very long Phase 8 talk-script example all pad the ~585-line body. Not a 2 because there is little explanation of things Claude already knows; not a 4 because several sections clearly could be trimmed. | 3 / 5 |
Actionability | Concrete and largely executable: a copy-paste beamer template, exact bash commands ('latexmk -$ENGINE -interaction=nonstopmode main.tex', 'pdfinfo slides/main.pdf | grep Pages'), state JSON format, and a parameter table. Minor gaps: 'generate_pptx.py' is described as a spec rather than provided code, and the 'spawn_agent' reviewer block is a pseudo-format rather than a real tool invocation. | 4 / 5 |
Workflow Clarity | Eight clearly sequenced phases with explicit validation and feedback loops: prerequisite checks ('which pdflatex && which latexmk'), paper existence checks, CJK detection switching ENGINE, a hard '⛔ STOP HERE and wait for user response' checkpoint, a compile error-handling loop (parse log → fix → recompile, max 3 attempts), page-count verification against the outline, and state persistence per phase. | 5 / 5 |
Progressive Disclosure | The body has clear section headers, but this is a ~585-line monolith with no bundle files: the full beamer template, talk-type slide tables, venue color table, and entire talk-script example are inlined rather than split into references. The single reference '../shared-references/reviewer-backends.md' points outside the bundle and is not clearly signaled. Structured but not well split — matches the 'content that should be separate is inline' anchor. | 3 / 5 |
Total | 15 / 20 Passed |