Content
77%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.
Highly actionable and exceptionally well-sequenced, with real validation feedback loops and domain-specific gotchas Claude would not know. The costs are token weight from repeated invariants and a monolithic structure that inlines reference material while simultaneously depending on external shared files, contradicting its own "zero external dependencies" rationale.
Suggestions
Move the style-preset tables (Phase 2 Beamer fix patterns, Phase 3c PPTX pitfalls, inspector JSON schema) into references/ files (e.g., references/fix-catalogs.md, references/style-presets.md), keeping SKILL.md as a phase-by-phase overview with clearly signaled one-level-deep links.
State each hard invariant once (in Hard Invariants) and reference it from Constants and the phases instead of repeating the reference-required, xhigh-reasoning, anonymity, and page-number rules verbatim in multiple sections.
Trim the "Empirical Origin" and "Prior Skill Relationship" sections to one or two lines each; the empirical narrative duplicates points already made in the Phase 1 rationale.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The operational core (commands, font tables, fix catalogs) is dense and non-padding, but the ~570-line body restates load-bearing rules multiple times (reference-required, xhigh reasoning, anonymity placeholders, page-number cap each appear 2-3 times across Constants, phases, and Hard Invariants) and includes narrative sections ("Empirical Origin", "Prior Skill Relationship") that could be cut or tightened. | 3 / 5 |
Actionability | Copy-paste-ready throughout: exact render commands ("pdftoppm -r 150 -png <pdf> <out-stem>", "soffice --headless --convert-to pdf"), full per-page Codex prompt templates, an executable find_shape Python helper with abort semantics, concrete LaTeX macros, and specific hex/pt values. The inspector and apply_fix are specified as contracts rather than code, with the flexibility explicitly justified ("Implementations may evolve; the contract is what downstream phases depend on"). | 5 / 5 |
Workflow Clarity | Phases 0-5 are clearly sequenced with explicit validation checkpoints: re-triage with a loop-back rule ("If > 2 NEEDS-WORK remain, loop back to Phase 1"), speaker-notes byte-level integrity check, anonymity fail-closed scan, assert-unique shape matching, checkpoint snapshots every 3 slides, and resumable state via POLISH_STATE.json — matching the feedback-loop-heavy top anchor. | 5 / 5 |
Progressive Disclosure | No references/, scripts/, or assets/ directories exist; everything (style-preset tables, Beamer/PPTX fix catalogs, inspector JSON schema) is inlined in one ~570-line SKILL.md, while six ../shared-references/*.md files are referenced from outside the bundle. Headers make it navigable, but the catalog and preset material clearly belongs in separate one-level-deep reference files. | 3 / 5 |
Total | 16 / 20 Passed |