The body is lean — a tight definition, a branch-selection guide, and six packed rules — with no padding and no explanation of concepts Claude already knows; every line earns its place.

Guidance is concrete and specific (naming conventions, runner examples like 'pnpm <name>', scratch-DB 'PROTOTYPE — wipe me', throwaway-branch capture), but it stops short of copy-paste-ready commands/code and omits key execution details such as the actual commit/branch command.

A clear sequence is laid out — identify the question, pick a branch with an explicit unreachable-user fallback, apply shared rules, then capture the verdict — providing decision feedback loops that function as checkpoints for a non-destructive process.

Sections are well-organized with clearly signaled one-level-deep links to LOGIC.md and UI.md, but those referenced files are absent from the bundle (no references/scripts/assets dirs), so the navigation leads nowhere and cannot earn a 3.

Names the domain (building one-off prototypes) and the two sub-contexts (state model/logic validation, UI exploration), but does not enumerate multiple distinct concrete actions the way a score-3 anchor like 'extract, fill, merge' does.

It states both what ('构建一次性原型来回答一个设计问题') and an explicit when trigger ('适用于...时'), matching the 'what AND when with explicit triggers' pattern that earns a 3.

The '适用于...验证某个 state model 或 logic...或探索 UI' clause supplies fairly natural triggers, but coverage of common phrasings is limited and leans on internal wording ('感觉对'), so it is not the broad natural-term coverage of a 3.

It carves a clear, narrow niche — throwaway prototypes for design-question validation via logic/state or UI exploration — with distinct triggers unlikely to fire for the wrong skill.