Verify UI

/verify-ui — Multi-modal UI verification

/verify-ui "<intent>" captures before/after screenshots and asks Claude vision to judge whether the visible change matches the stated intent — and whether anything else changed unexpectedly.

Pattern from Mano-verify (arXiv 2509.17336, Sep 2025) — the verifier reasons over (pre-image, post-image, intent text) jointly, not just text. #1 on OSWorld specialized at 58.2%.

When to use

• Pixel-shifted layouts (margins, alignment)
• Color or typography drift
• Component visibility (something appeared/disappeared)
• Unintentional regressions (you fixed component A, but B looks broken now)
• Accessibility-relevant visual changes (focus rings, contrast)

When NOT to use

• Pure logic changes (use /verify with the verifier agent — runs tests, no UI)
• Backend API changes
• When the dev server isn't running and you can't start it

Protocol you must follow

1. Confirm intent

• The user told you what they wanted ("change the button color to blue", "add a loading spinner", …). Restate it back in 1 sentence.
• If intent is vague ("make it look better"), STOP — ask for specificity. UI verification needs a concrete claim.

2. Capture BEFORE state

• Find the dev server config: .claude/launch.json or ask the user
• preview_start if not running
• preview_screenshot — save the result. This is the BASELINE.

3. Apply the change

• Edit/Write the actual code
• Reload the preview (preview_eval with location.reload() or wait for HMR)
• For visible-immediately changes, capture stops here

4. Capture AFTER state

• preview_screenshot again. This is the POST-CHANGE state.

5. Vision judge

You have BEFORE and AFTER images in context plus the intent text. Reason explicitly:

• Did the intended change occur? (Compare before and after for the specific feature mentioned in intent.)
• Did anything else change? (Scan the rest of the screenshot for layout shifts, color drift, missing elements, broken icons.)
• Is there an obvious regression? (Cut-off text, overlapping elements, broken hover states, wrong-direction shifts.)

Don't pixel-diff — diff meaningfully. Anti-aliasing, font hinting, browser version drift produce pixel changes that aren't real regressions.

6. Report

INTENT: <one sentence>
BEFORE: <path to screenshot 1>
AFTER:  <path to screenshot 2>

VERDICT: pass | fail | escalate
CONFIDENCE: 0.0 - 1.0
INTENDED_CHANGE: <observed | not observed | partial>
UNEXPECTED_CHANGES: <none | list with descriptions>

Last line: RESULT_verify-ui=PASS|FAIL|ESCALATE

ESCALATE = "I see changes I can't classify as expected or regression — surface to lead."

Use with /aspect-panel

For high-stakes frontend work (release-blocking, accessibility-critical, brand-affecting), run BOTH:

1. /verify-ui — visual judge of the rendered change
2. /aspect-panel with accessibility-reviewer agent — code-level a11y check

Together they catch:

• visual drift (panel can't see)
• a11y regressions in code (visual judge can't see semantic structure)

DOM + screenshot dual-grounding (advanced)

Going beyond pure pixel comparison: preview_inspect returns computed styles + bounding box for any selector. Use this when the user's intent is precise:

• Intent: "make the button 8px taller"
• preview_inspect on the button selector — read computed height before
• Apply change
• preview_inspect again — confirm height is +8px
• Then screenshot to confirm no other layout shift

This is more reliable than pure vision for measurable changes.

Anti-patterns

• Asking the user "does this look right?" — that's not verification, that's deferral. Capture the screenshots and judge yourself; surface low confidence as ESCALATE.
• Pixel-diff with no semantic understanding — vision-LLMs are better than diff tools at ignoring noise (anti-aliasing, sub-pixel rendering). Use them.
• Skipping the BEFORE capture — without baseline, you can't claim regression-free. Capture even if it costs an extra round-trip.
• Verifying with the dev server in a stale state — always reload between change and AFTER screenshot.
• Running on a 1-line CSS color change — overkill. The dev server preview shows it instantly; you don't need a Mano-verify pass.

Cost

• 2 screenshots × ~50ms each (cheap)
• 1 vision-LLM call (~$0.01-0.03 per check at Claude pricing)
• Compare to Applitools enterprise ($1k/mo) or running full Playwright visual regression suite (minutes per change)