# UI Sweep > Audits shipped examples in a real browser and fixes interaction and visual defects. Use after runtime or theme changes, or as a standing dispatch. - Skill: `max-sixty/ui-sweep` (Agent Skill) - Install (CLI): `npx skillmds@latest add max-sixty/ui-sweep` - Raw SKILL.md: https://api.skillmd.com/api/skills/max-sixty/ui-sweep/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: max-sixty (https://skillmd.com/u/max-sixty) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/max-sixty/ui-sweep --- # UI sweep Use the gallery and a composed example to find defects beyond the suite's stated invariants. Derive experiments from the reader's task and the code's interaction structure, then drive, judge, and fix them in the browser. ## First reading Open the exact candidate before reading its implementation or the author's verdict. Use the page for its stated task at ordinary browser zoom. Record the initial visual problems in the subject, actions, evidence, and supporting information. Give an independent reviewer the task and candidate before giving them the author's assessment. ## Derive Read the changed code and its callers. For a standing sweep, start from a reader workflow in the examples. Use `skills/leaf/assets/CLAUDE.md` for UI contracts, render gates, and the ownership map; read the relevant owners' module headers. Source access belongs in this audit, including when delegating it. Keep a short working interaction model: the reader's goal; the source, control, surface, focus, and retained state involved; and the transitions that create, transfer, or end their relationships. Identify where different owners cooperate. Derive expected behavior from the contracts and reader's goal; use implementation to establish reachable states and mechanisms. Name unsettled design assumptions. ## Challenge Before validating a proposed result, write competing failure hypotheses for the riskiest relationships. Give each a mechanism, an observable consequence, and an experiment that could disprove it. Prioritize changed ownership boundaries and relationships the existing tests leave unexercised. Generate experiments by perturbing the model: nest or replace an interaction, reverse or cancel it, exhaust its available room, or change external state while it remains active. Choose the operators that challenge the hypothesized mechanism. Vary one factor at a time across an ownership boundary and exercise both the forward and return paths. State what must change and what must survive before driving. When a change sends an existing surface into a different posture or fallback, compare the same reader task against the merge base; the candidate must preserve its task-relevant information and actions. Include an unrelated change that should preserve the active relationship. Choose content density, viewport boundaries, input routes, and visual states that distinguish the hypotheses. Carry a tested relationship into a different composed example to challenge assumptions the gallery makes easy. For a changed layout, find a width or content length where a group first wraps or rearranges and inspect both sides of that transition. Include a constrained height when the surface claims to fit the viewport. ## Observe Re-vendor before testing. Use `serve` and `open_page` from `tests/render_harness.py`, or `scripts/preview.py --automation`. These process-owned servers exercise the real HTTP and event-log loop without delivering reader feedback to the task. Give independent runs separate page state. Drive with real input and confirm each intended transition occurred. For each experiment record initial state → input → expected relationship → observed result, with the evidence that decides the claim. Read focus, surface and owner identities, and retained values where those matter. Measure geometry for spatial claims; inspect paired screenshots for hierarchy, spacing, and paint. For motion, capture the path as well as endpoints: `tests/CLAUDE.md`, "Distinguish a frame, a sequence, and an instant", owns the recording and `HOLD_MOTION` mechanics. The hold patches `Element.prototype.animate`; CSS animations bypass it, so sample their geometry frame by frame. Settled boxes cannot establish a smooth transition. For paint-only transitions, compare the same target's hit-test rectangle before, during, and after the state change; added paint does not move its target. Judge both the whole composition and its details at native scale. Inspect the subject, actions, evidence, and supporting information; quality in one region says nothing about another. Check whether labels and values group and align, text wraps at meaningful boundaries, controls and icons remain legible, and space follows the active task. Capture the full viewport for composition and native-scale details wherever the full image cannot support that judgment. Computed styles and geometry explain a visual result; they do not establish that it looks coherent. Test affected paint in both color schemes and affected export behavior in print. Use visual treatments to communicate content hierarchy and state. Avoid rounded one-sided borders and reflexive cards, tints, gradients, or soft shadows. Keep visual grammar in browser judgment; automate the concrete failures it causes. ## Reconcile Give every finding a disposition, including discoveries outside the planned route: reproduced and fixed, disproved with evidence, or unresolved with the missing evidence named. Fix reproduced defects at their owning boundary, then repeat the failing experiment and a sibling derived from the same relationship. Pin measurable page defects in render_version and gesture defects in the owning `test_render_*.py` module. Put the bug back once and watch the new check fail. Keep the implementation, tests, and owning contracts aligned, and run the required checks. Report established behavior, unresolved findings, and untested relationships explicitly; uncompleted transitions remain untested. Report functional and visual coverage separately, naming the regions and states judged in the browser. Passing interaction checks or collecting screenshots is not a visual verdict. Hand over the reviewed branch and evidence; landing waits for the go-ahead.