Design critic
| ID | ap-design-critic |
| Kind | subagent |
| Selection | optional |
| Category | Design / Critique |
| Targets | claude, codex, cursor |
Independently critique rendered visual design for product fit, identity, composition, typography, hierarchy, distinctiveness, restraint, and craft.
When to use
Section titled “When to use”Use when you want independent, read-only visual judgment of rendered screenshots or frames against a clear brief.
How it is selected
Section titled “How it is selected”Optional. Select during initialization or when offered in an update, or install separately when you need it.
Install
Section titled “Install”agents-pack install ap-design-critic --dry-runagents-pack install ap-design-critic --yesHow to invoke
Section titled “How to invoke”Launch a fresh ap-design-critic with no inherited conversation. Give it theattached desktop and mobile screenshots and this brief: a calm researchlibrary for scientists, with readable abstracts and clear search hierarchy.Ask for the highest-impact visual improvements and what to preserve.What it reviews
Section titled “What it reviews”The critic is a read-only subagent that judges supplied visual artifacts against the product objective, audience, constraints, and optional references. It does not inspect source code, CSS, design tokens, implementation rationale, or repository history to justify the design.
Its scorecard covers product fit, design identity, composition, typography, visual hierarchy, distinctiveness, restraint, and craft. An overall studio-quality score expresses its judgment of the whole design. The report also identifies what works, the biggest gaps, three prioritized improvements, and details to preserve.
Typography judgment distinguishes intentional, readable overlap from clipped glyphs, accidental collisions, and text blocks that almost touch. It also checks whether weight creates hierarchy or whether too many headings, labels, buttons, and metadata lines are bold and competing for attention.
Freshness is required in a refinement loop
Section titled “Freshness is required in a refinement loop”Every assessment needs a new conversation without inherited history. Earlier designs, critiques, scores, targets, and the implementing agent’s rationale must not be supplied. A new agent ID with copied history is insufficient; asking an existing critic to forget its context does not restore independence.
If contaminated context is already present, the critic requests a clean dispatch and does not issue a score. If it cannot inspect the supplied visuals, it reports that limitation rather than inferring visual quality from text.
You can explicitly request a standalone comparison of several supplied variants. That comparison can recommend a winner or combinations, but it is a different task from a fresh assessment inside a refinement loop.
Visual judgment and functional verification
Section titled “Visual judgment and functional verification”Scores are subjective guidance, not proof of correctness or an automatic choice for the user. Pair the critic with frontend review to check responsiveness, interaction states, accessibility, and runtime behavior.