AGENT SKILL

design-quality

Judge rendered UI against an approved design and resolve visual critique before final evidence.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
design-quality
description
Judge rendered UI against an approved design and resolve visual critique before final evidence.

02 资源文件 3

03 SKILL.md 正文

For direction exploration or an approved design handoff, load design-direction once and use its current project brief. Reuse that context and selected phase.

Let Impeccable own the design workflow. Load its current skill, run context once at the actual app path, and choose its one matching playbook. This skill adds the acceptance contract; it is not another router or another review loop. Reuse the active execution method when assigning the critic; its actual vision capability and the selected provider restrictions govern the assignment.

  1. Identify the surface, user task and incumbent visual language. For unresolved direction, use design-direction; read direction.md only when more exploration guidance is needed. For refinement, preserve approved identity and select references only for the gap at hand.
  2. Before editing, read Impeccable's craft floor. Name the concrete result: hierarchy, density, type, assets, responsive behavior and interaction states. Record approved choices and explicit exclusions in the surface brief.
  3. Implement the selected direction. Use real assets when their medium matters; follow assets.md. Reuse the product's components and tokens. Keep operational UI focused on the user's next action. For a new shared visual language, compare one representative screen early before propagating its composition. This is an internal implementation check, not another user approval. Use matching viewport, state and representative content; distinguish real empty-state constraints from missing fixture data.
  4. Complete Impeccable's selected workflow, including its finish checks. Have the orchestrator assign an independent visual-capable critic current working screenshots, task and usage context, confirmed constraints, approved direction when one exists, and visual references with the function of each reference. Do not give that critic implementation rationalizations, code effort, prior critique or self-scores. A distinct finish reviewer can satisfy this role; a designer's self-review cannot.
  5. The critic reports separately (a) conformity with confirmed constraints and (b) compositional quality against the reference level, then returns specific visible gaps ordered by impact. A grounded score may support either judgment; it never replaces criteria, proves approval or triggers repetition by itself. Correct in one batch, then have the critic confirm changed states. Run one or two initial passes and evaluate convergence. Extend only for named unresolved gaps or newly observed regressions. If the same hypothesis is not improving, change it instead of repeating the loop. Judge the approved task hierarchy, density, grouping and content area as well as responsive geometry. A material mismatch with an approved choice blocks visual acceptance; it cannot be dismissed as optional polish because nothing clips. Return the compared criteria, deviations and review scope. Closing a bounded finding list proves that list only. The parent checks the complete requested visual outcome before declaring acceptance. “Ready to compare” and “no defects found” within the review scope do not establish user selection.
  6. Once critique findings are resolved, prepare the final delivery evidence through evidence-capture from that reviewed candidate. Working screenshots and functional probes used to review are diagnostic material, not the final evidence package. A subsequent visual change reopens review only for the affected surface. Do not present unfinished design evidence as ready.

Finish with what changed visually, the compared evidence, unresolved gaps and what the user can now do. Scanner success proves mechanical checks only. A full functional check does not certify composition. Readable empty, loading, error, populated and constrained states matter more than the count of captures.

See sources.md for the adopted ideas and their limits.

04 同仓库更多 Skills