Content
73%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 is an exceptionally actionable, rigorously validated workflow document: enforced ordering, feedback loops, exact measurements, and executable code throughout. Its main liability is size — ~39k tokens with signature tables and a findings catalog inlined that belong in the existing reference files, plus discursive prose that repeats rules across sections.
Suggestions
Move the §4 helper-signature tables and the §12 findings catalog into existing or new reference files (e.g. an api.md and findings.md beside charts.md), keeping one-line pointers with the existing 'open it when...' signals.
Tighten the multi-paragraph narratives (e.g. §3.5's caption/footer geometry and the §3 blank-region rules that recur in §12) to the rule plus one worked example, cutting the measured back-stories.
Either surface the five layouts-*.md sub-references from SKILL.md directly or fold their pointers into the layouts.md entry so no content sits two reference levels deep.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body runs ~2130 lines (~39k tokens): §4 inlines two full helper-signature tables ('Read this rather than the modules (51k tokens of source...)') while charts.md/shapes.md exist as separate files, §3.5 spends ~60 lines on caption/footer geometry including measured back-stories ('on five reference pages, three of them at 6.40in and 6.44in'), and rules recur across §3, §6.5 and §12. This is 'noticeably verbose; several unnecessary explanations or padded sections' — not 1 because it never explains basic concepts Claude already knows and nearly every passage is operational; not 3 because the sheer inlined API/ findings bulk goes beyond 'could be tightened'. | 2 / 5 |
Actionability | Fully executable and copy-paste ready: a complete program from imports to 'prs.save(os.environ["PPT_OUTPUT"]) # the last line of every program', exact thresholds ('body copy stays at or above 14pt', 'the median is 250'), executable geometry ('assert stack(right).short_by(*PLAN, *((GUTTER,) * (len(PLAN) - 1))) == 0.0'), and per-failure remediation ('Under 1 is refused, because a cover that does not cover is fit="contain"'). Specific examples cover the common cases. | 5 / 5 |
Workflow Clarity | The sequence is explicitly enforced, not suggested — 'ppt_build refuses until ppt_prepare has read the task', 'ppt_outline records it, and the build refuses until it exists' — with validation checkpoints at every stage and a full refusal/report catalog in §12 naming the move for each. Feedback loops are explicit: 'Fix everything a build reported before you build again, and send the edits and the build in one reply' and 'Measure every band, ask short_by, then draw' give validate-fix-retry cycles for error recovery. | 5 / 5 |
Progressive Disclosure | Seven reference files are linked with purpose signals ('Open it now, before the first page', 'Open it when a page carries a table', 'Open it before choosing a form') and all seven exist in references/ (charts.md, design-requirements.md, formulas.md, icons.md, layouts.md, shapes.md, tables.md). Not 5: layouts.md itself links to five further files (layouts-data/figures/multiples/primitives/type.md), making those two levels deep and undiscoverable from SKILL.md, and the inlined helper API and findings catalogs are content that could live in the existing references. | 4 / 5 |
Total | 16 / 20 Passed |