| name | ui-iterate |
| description | After editing a UI file (.tsx/.jsx/.css/.vue/html) with a dev server running: screenshot 375/768/1440px, critique hierarchy, spacing and states, fix and re-shoot. |
ui-iterate — look at what you built before calling it done
UI edited blind is UI shipped broken. Code that compiles says nothing about what renders.
This loop forces the screenshot-critique-fix cycle that separates polished frontends from
plausible ones. Run it after every meaningful UI change, before declaring the task done.
The loop
-
Find the live page. Locate the dev server: check for a running process on common ports
(3000, 5173, 4321, 8080), $CONDUCTOR_PORT, or the dev script in package.json (start it
if the repo has one and nothing is running). Identify the route your change affects.
-
Screenshot three widths. Using the browser tools (claude-in-chrome; resize_window
then computer screenshot), capture the affected page at:
- 375px — phone
- 768px — tablet
- 1440px — desktop
If browser tools are unavailable, use a headless fallback (
npx playwright screenshot or
agent-browser) — never skip the looking step because the preferred tool is missing.
-
Critique what you see. Judge the screenshots, not the code, against:
- Hierarchy — is the most important element visually dominant? Does the eye land where
the page's job says it should?
- Spacing — consistent scale (4/8px grid), no cramped clusters or orphaned gaps,
breathing room around groups.
- Typography — clear size/weight contrast between levels; line length 45–75ch; no
near-identical sizes competing.
- Contrast & accessibility — text meets WCAG AA against its actual background;
interactive elements look interactive.
- States — hover, focus, empty, loading, error, long-content overflow. Trigger the ones
your change touches and screenshot them.
- Responsive integrity — nothing clipped, wrapped awkwardly, or overlapping at any of
the three widths.
-
Fix the worst thing, re-shoot. One issue at a time, most severe first. Re-screenshot
the affected width after each fix. Repeat until step 3 finds no majors.
-
Measure the change, do not eyeball it. Comparing two screenshots by looking at them
misses exactly the regressions that matter: the 4px shift on a width you were not focused
on. Keep the pre-change capture as a baseline and diff against it:
npx agent-browser diff screenshot --baseline .context/ui-evidence/before-1440.png \
--threshold 0.02 -o .context/ui-evidence/diff-1440.png
It reports a mismatch percentage and writes a PNG with the changed pixels highlighted. Run
it on the widths you did not intend to change — a non-zero mismatch there is an unintended
regression, and that is the whole point. diff snapshot compares the accessibility tree
instead, which is less flaky than pixels for structural changes.
Two things make pixel diffs trustworthy rather than noisy: a fixed viewport
(agent-browser set viewport 1440 900) and waiting for the network to settle
(agent-browser wait --load networkidle) before every capture. Without both, you are
diffing animation frames.
-
Show the evidence. The final message includes what the screenshots showed, the mismatch
percentages, and what changed because of them. For gated repos, save the final captures to
.context/ui-evidence/ so verify.sh or a reviewer can see the loop actually ran.
When NOT to use
Pure logic/data changes with no rendered surface, or repos with no runnable frontend. Do not
fake the loop with code-reading — if nothing can render, say so plainly instead.
