Content
92%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.
An unusually disciplined operational skill: extreme token efficiency, fully executable commands and file shapes, and two numbered workflows with an explicit check-until-zero validation loop and error-recovery paths. The weak point is progressive disclosure: the ten referenced reference pages are missing from the bundle as provided, and the inlined canvas.json/notes/sharing spec plus the cross-skill substitution table add navigation friction that keeps it at the midpoint.
Suggestions
Ship the ten referenced pages (format.md, craft.md, diagrams.md, print.md, questions.md, view-state.md, design-system-components.md, brand-colors.md, brand-typography.md, make-from-this.md) under artifact-type/reference/ in the skill folder, or rewrite the body to be self-contained — as bundled, every 'read X FIRST' instruction points at a file that does not exist.
Move the notes-shape, pages/launch and Sharing link-role detail into a reference page and keep SKILL.md to the core create/revise loop plus command reference; the canvas.json paragraph is currently a several-hundred-word inline schema.
Replace the 'Reading the references in a canvas' substitution table by editing the reference pages to use this skill's own terms (project/<path>, notes, <root>), so readers do not have to apply a mapping table mid-task.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is telegraphically dense with zero conceptual filler — it never explains what a design system or canvas is, compresses rules to their minimum ("Write only under `<root>/project/`: a file elsewhere never shows", "NEVER two artboards in one message"), and pushes all detail to reference files. Apparent repetition (the `check` loop, "NEVER VERIFY", "never retype") is contextual restatement of critical constraints in the Creating vs Revising workflows rather than padding, so it matches anchor 5's 'every token earns its place'. | 5 / 5 |
Actionability | Guidance is copy-paste executable throughout: an exact CLI surface with signatures and outputs ("`upload <root> <file>` → `{url:"/_blob/<id>", …}`", "`check <root>` → `problems` (0: done)"), a complete, runnable `.dc.html` skeleton, the full `canvas.json` literal with every key's shape, and precise fallback rules ("Can't upload… an image becomes a labelled placeholder; say so"). This matches anchor 5's fully executable, common-case-covering example. | 5 / 5 |
Workflow Clarity | Creating (steps 1–6) and Revising (steps 1–4) are explicitly sequenced with an embedded validate-fix-retry feedback loop — "Then run `check` and fix what it names until `problems` is 0" (repeated in Revising step 4) — plus recovery paths for refused uploads, mid-run parallel-edit conflicts ("read right before you edit"), and unanswerable questions ("Never end on a question nobody can answer: decide and build"). This is anchor 5's clear sequence with explicit validation and error-recovery loops; the only nit is that 'The calls' command reference is defined after Creating first references it. | 5 / 5 |
Progressive Disclosure | References are well signaled and one level deep — each named with its read trigger ("`format.md` (before a first artboard)", "`print.md` (FIRST for print)", "`make-from-this.md` FIRST") — but the referenced markdown files (format.md, craft.md, diagrams.md, print.md, questions.md, view-state.md, design-system-components.md, brand-colors.md, brand-typography.md) do not exist in the provided bundle, where `artifact-type/reference/` holds only Google-Fonts JSON data. Combined with a dense inlined spec (full canvas.json schema, notes/sharing rules) and a substitution-table section ('Reading the references in a canvas') that makes readers translate another skill's vocabulary, structure is present but imperfect — between anchors 3 and 4 rather than the 'minor organization gaps' of a clean 4. | 3 / 5 |
Total | 18 / 20 Passed |