Lint Your Screens with a Multimodal Pre-Critique

Workflow area: Ideation & Critique · Tier: Foundation · Hands-on time: ~25 minutes

Half of every critique session gets spent on findings that needed no taste to find: the truncated label, the orphaned alignment, the button that reads "Submit" next to one that reads "Save changes." Those are lint. A multimodal model with a rubric catches most of them in two minutes, which changes what critique is for — the room stops proofreading and starts designing. The trap to avoid is treating the model's opinions about your design direction as findings. The rubric exists to keep it on lint duty.

When to use — and when not to

• Use it before every human critiquearriving pre-linted is respect for the room's time and buys you deeper feedback.
• Use it before stakeholder reviewsa VP who spots a typo stops trusting the strategy; lint protects the conversation you actually need.
• Use it on AI-generated frames especiallygenerated UI has its own defect signature (fake data, near-miss alignment, invented icons) that this pass catches cheaply.
• Skip it for napkin-fidelity worklinting a sketch punishes exactly the roughness that makes early sharing safe.
• Never let it replace human critiquethe model has no idea what your product strategy is, who your users are, or why this screen exists; it lints artifacts, it cannot judge intent.

The recipe

1. 1.

   Export clean screenshots at real size. One screen per image, realistic content in place (pb-07's states count as screens). Vision models miss less at higher fidelity, and lorem-ipsum screens hide the copy findings entirely.

2. 2.

   Run the rubric-anchored lint. In Claude, ChatGPT or Gemini, attach the screenshots and run:

   Copy

   You are a design linter. Review these screens against ONLY this
   rubric. Do not comment on design direction, aesthetics, or ideas.
   
   Context: [ONE LINE — product, platform, this screen's job]
   
   Rubric:
   1. HIERARCHY — is the primary action visually primary? Do sizes/
      weights match importance? Anything competing that shouldn't?
   2. ALIGNMENT & SPACING — misalignments, inconsistent gaps, elements
      off-grid.
   3. CONTRAST & LEGIBILITY — text/background pairs that look
      insufficient, type too small for the platform, truncation.
   4. COPY — inconsistent terminology across screens, unclear labels,
      case inconsistencies, buttons whose labels don't say what
      happens, placeholder text left in.
   5. CONSISTENCY — same element styled differently across these
      screens; interaction patterns that contradict each other.
   
   For each finding: SCREEN | RUBRIC ITEM | WHAT & WHERE | SEVERITY
   (fix-before-sharing / fix-before-shipping / judgment-call).
   If a rubric category has no findings, say "clean". Do not pad.
   

3. 3.

   Triage in one pass — fix, flag or discard. Good lint output looks like this; note the third line is the model overstepping, and you discard it:

   Copy

   Screen 2 | COPY | Primary button "Submit" vs. Screen 4 "Save
   changes" for the same action | fix-before-sharing
   Screen 3 | CONTRAST | Secondary text on tinted card looks below
   AA at body size | fix-before-shipping (verify with a real checker)
   Screen 1 | HIERARCHY | "Consider a card-based layout instead" |
   DISCARD — direction opinion, outside rubric
   

   Mechanical findings you fix immediately. Contrast findings get verified with a deterministic checker — the model estimates contrast from pixels; Stark or a WCAG checker computes it, and only the computed number counts (pb-31 owns the full pass). Direction opinions get discarded without guilt.

4. 4.

   Bring judgment-calls to the humans. Findings the model marked judgment-call — plus everything it can't see: strategy fit, user context, emotional register — are your critique agenda. Open the session with "it's pre-linted; here's what I actually need eyes on."

5. 5.

   Feed repeat findings upstream. The same lint appearing three weeks running is a system gap, not a you gap — a missing component variant, an unwritten copy rule. Route it to pb-27's voice system or your design-system backlog, and the lint stops recurring.

Make it yours

• Claude for copy-heavy screens: consistently strong at terminology drift and label-versus-behavior mismatches across a screen set — lead with it when the surface is dense with words.
• Gemini or ChatGPT as a second linter on high-stakes reviews: different vision models miss different defects; for a launch review, run two and union the findings — the overlap is also a confidence signal.
• Team rubric as a shared asset: ratify the rubric wording with your team (add your design system's specific rules — token names, voice rules from pb-27) and store it where everyone prompts from the same one; a shared linter beats five private ones.

Quality check

• Every finding traces to a visible pixelthe model must say where; a finding you can't locate on the screen is a hallucinated finding, and vision models do invent UI elements.
• Direction opinions were discarded, not absorbedthe rubric boundary held; if "consider a different layout" survives triage, the linter has become an anchoring critic.
• Contrast claims were re-verified deterministicallya model's contrast eyeball is a hypothesis, never a compliance result.
• "Clean" categories appeared sometimesa linter that always finds something in every category is padding to please you.
• Human critique still happenedthe pre-critique's success measure is better human sessions, not fewer of them.