Content
90%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 exemplary lean, highly actionable rulebook: exact tool names, commands, parameters, and real validation/feedback loops for destructive and batch operations, with details correctly pushed to one-level-deep reference files. The main gaps are the implicit (section-ordered rather than explicit) workflow sequencing and a dangling cross-skill reference to ../common-ui-design/references/human-direction.md.
Suggestions
Fix or inline the broken cross-skill link [Human Direction](../common-ui-design/references/human-direction.md) (referenced in both the body and prompting.md) — either ship the file in the bundle or summarize the keep/change/avoid essentials locally.
Make the end-to-end sequence explicit (e.g., a short numbered flow: preflight → read/audit → confirm direction & approval → variant/write → lint → re-audit) so readers don't have to infer order from section arrangement.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, imperative bullets with zero concept explanations — e.g., "Never overwrite originals. Fix through `generate_variants` (default `creativeRange: REFINE`, `variantCount: 1`)" — assuming Claude's competence throughout. Every token earns its place; nothing is padded, matching the lean score-5 anchor rather than the minor-trimming case at 4. | 5 / 5 |
Actionability | Guidance is fully executable for an instruction-only skill: exact tool names (`generate_variants`, `apply_design_system` takes screen instance ids from `get_project`), runnable commands (`npx @google/design.md lint DESIGN.md`, `node scripts/audit_html.js <dir>`), and concrete parameters ("poll `get_screen` every 30 s, max 10", "longest edge <=1600 px"). This matches the copy-paste-ready score-5 anchor; nothing is pseudocode or vague. | 5 / 5 |
Workflow Clarity | Sections run in execution order (Preflight → Read → Human Direction → Write Safely → DESIGN.md → Prompting) with real validation checkpoints and feedback loops — "must report 0 errors and 0 warnings before upload", "re-audit after each pass; max 2 `edit_screens` passes", "On timeout do not retry; poll", and approval gates before destructive ops (so the destructive/batch cap does not apply). It falls short of 5 because the end-to-end sequence is implied by section order rather than an explicit step ordering, and some checkpoints (screenshot verification, bounded review) delegate procedure to references. | 4 / 5 |
Progressive Disclosure | The SKILL.md is a genuine overview with well-signaled, verified one-level-deep references — [Stitch MCP](references/stitch-mcp.md#bounded-image-review), [Prompting](references/prompting.md), [DESIGN.md](references/design-md.md), plus a bundled script — all of which exist with matching anchors. It misses 5 because [Human Direction](../common-ui-design/references/human-direction.md) is referenced twice (body and prompting.md) but does not exist in this bundle, a navigation dead-end for readers. | 4 / 5 |
Total | 18 / 20 Passed |