verify-like-a-user
Skill1 starsChanged 5 months agoMITView on GitHub
Build and check UI the way a person meets it — measure the rendered page instead of trusting the source, break it on purpose at several widths and along the paths people actually take, fix the root cause, and re-measure before calling it done. Carries a catalogue of UI failures that survive code review (clipped borders, dropped size classes, dead ends on short screens, stale caches, redirects that never move the URL) with symptom → cause → fix. Triggers on "verify this UI", "does this look right", "check it at other widths", "it looks bad, fix it", "break the product", "make these pages consistent", "why is this clipped / overflowing / huge / not clickable", "audit every page in this section", and any screenshot sent with "fix this".
In this repository51 other files it gives its agents.
- architect-enterprise-aiSkill
- author-agent-skillsSkill
- build-agent-harnessesSkill
- build-aiewf-2024-ragSkill
- build-generative-mediaSkill
- build-multimodal-productsSkill
- build-rag-search-stacksSkill
- build-robotics-aiSkill
- build-with-deepmindSkill
- build-with-microsoft-aiSkill
- build-with-openaiSkill
- deploy-public-sector-aiSkill
- design-agent-skillsSkill
- evaluate-with-braintrustSkill
- lead-ai-engineering-teamsSkill
- lead-ai-transformationSkill
- learn-summit-2023-remoteSkill
- learn-summit-2023-talksSkill
- navigate-aie-europe-2026Skill
- navigate-aiewf-2024-keynotesSkill
- observe-ai-productionSkill
- operate-agent-memorySkill
- operate-legal-ai-agentsSkill
- operate-openclaw-agentsSkill
- optimize-llm-inferenceSkill
- plan-sovereign-aiSkill
- price-ai-productsSkill
- run-aiewf-2024-evalsSkill
- run-llm-evalsSkill
- run-summit-2023-workshopsSkill
- secure-ai-agentsSkill
- ship-ai-as-pmSkill
- ship-aiewf-2024-agentsSkill
- train-llms-locallySkill
- architecture-reviewSkill
- chat-failure-auditSkill
- create-a-skillSkill
- cross-model-reviewSkill
- cto-weekly-reviewSkill
- friction-auditSkill
- ig-carouselSkill
- pdf-reportSkill
- raise-prSkill
- reel-editSkill
- research-reportSkill
- ship-checkSkill
- spec-grounded-designSkill
- state-lifetime-decisionSkill
- ste100Skill
- scroll-engineeringSkill
- ui-uxSkill
The file
We split this skill into the rules it gives an agent
11 rules. About 0 are common, 8 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 43
If the user's screenshot and your reading of the code disagree, the screenshot is right.
- Very specific: line 51
Read the project's design rules first: DESIGN.md, CLAUDE.md, the tokens file, globals.css.
- Distinctive: line 56
Read laws.md. Decide, before touching anything, which structure the screen needs: tabs vs rail, popup vs page, one column vs grid, one …
- Distinctive: line 60
Do not start from what the previous version did.
- Distinctive: line 118
Before writing a fix, look the symptom up in patterns.md.
- Distinctive: line 141
Never gate the input on decoration.
- Very specific: line 148
Run the exact evals from Step 2 again.
- Distinctive: line 152
Do not call a flow done from one width, one path, or one screenshot.
- Very specific: line 171
Write the record to {output_dir}/verify-YYYY-MM-DD-[slug].md in break-it and overhaul modes: defects ranked, each with what you did / …
- Distinctive: line 182
Search patterns.md before diagnosing from scratch.
- Distinctive: line 184
Never commit a test file you did not run after editing it.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Verify Like a User
You are not done when the code compiles. You are done when the page, as a person meets it, does what they asked — and you can show the measurement that says so.
This skill is the discipline that turned a week of "looks fine in the diff" into
shipped UI: ground truth first, measure not eyeball, break it on purpose, fix the cause
not the symptom, re-measure, report honestly. It pairs with ui-ux (which
audits and polishes) — this one is about the loop that makes any UI change true.
Supporting files — read the one the step names:
- laws.md — the UX laws, each with the concrete rule it produced
- patterns.md — failure catalogue: symptom → cause → fix (search it first)
- measure.md — DOM measurement recipes and browser-automation gotchas
Output: {SKILL_OUTPUT_DIR}/verify-like-a-user/ — see ../OUTPUT.md
Step 0 — Mode
| Mode | When | Ends with |
| fix-a-screenshot | user sends a screenshot and says what is wrong | a measured before/after on that exact element |
| verify-a-change | you just changed UI | measurements at four widths + the click-through path |
| break-it | "test like a user", "break the product" | a ranked defect list with proof, then fixes |
| overhaul | "this section is bad, redo it" | audit → structure decision → rebuild → re-verify every page |
If the user's screenshot and your reading of the code disagree, the screenshot is right. Find why the code you read is not what rendered (patterns.md § "The class is in the source but not on the page").
Step 1 — Ground truth before pixels
- Read the project's design rules first:
DESIGN.md,CLAUDE.md, the tokens file,globals.css. Note what a page title is (size/weight/face), what a section header is, which radius tier means what, which tokens are serif. - Find the page in the product that already does this right (the real settings page, the real list page). Copy its classes, not its idea.
- Read laws.md. Decide, before touching anything, which structure the screen needs: tabs vs rail, popup vs page, one column vs grid, one button that flips vs two.
Do not start from what the previous version did. Start from what the rules say and what the user said — in that order when they conflict, and say so.
Step 2 — Measure, don't eyeball
A screenshot tells you something is wrong. A measurement tells you what. Use
measure.md — every check below is one eval.
Minimum for any change to a screen:
- Widths: 1440, 1000, 640, 400. Primary action reachable at each (in the viewport, or the container scrolls to it).
- Rects: the two things that must not overlap, do not. The thing that must align with the box above it, does (same left/right edge to the pixel).
- Computed styles: the font-size and family that rendered on the element you care about, not the class you wrote. Border width and style. Border radius on a tab.
- Overflow: which ancestor scrolls, and whether the content is taller than it.
scrollHeight > clientHeighton the wrong element is the dead end. - Text:
scrollWidth > clientWidthon any title = clipping. Line-clamp actually clamping (2 lines, not 4). - State:
aria-selected,aria-pressed,aria-invalidon what you think is active.
Take the screenshot after the measurement, to see what the numbers mean.
Step 3 — Break it on purpose
Your own happy path proves nothing. Do these every time, and log each as pass or fail with the number that proves it:
- The click-through path, not the direct URL. Navigate from where a person would (the list, the sidebar, the card). Direct load and client navigation render differently; redirects, layouts and loading states differ between them.
- Reload and Back mid-flow. Is typed work lost silently? Is there a
beforeunloadguard where there should be? - Short window. 1000×560. Is the primary button below a fold nothing scrolls to?
- Empty and full. Zero items, one item, forty items. Does an empty shelf say "you have everything" or lie with "no matches"?
- Invalid input. Clear a required field, paste 10,000 characters, four emoji (JavaScript counts UTF-16 units, not graphemes). Does the error name the field?
- Rapid double-click on any button that creates something. Count what got created.
- Mid-animation frames. Trigger a close or a transition and screenshot at ~80ms. Seams, flashes and stray hairlines live there.
- Stale state. Change data another way (a script, another tab) and reload. Caches with hour lifetimes, and code that assumes a row still exists, fail here.
- A second machine's view. Different host (
localhostvs127.0.0.1), fresh session, no cookies. The dev server treats them differently.
Rank what you find by what a person loses: dead ends and lost work first, lies second ("Added" on something you cannot act on), looks last.
Step 4 — Fix the cause
Before writing a fix, look the symptom up in patterns.md. Most of these have been paid for already:
- a rendered size that ignores the class → the class merger dropped it
- a border cut off at one edge → it was a shadow outside the box, clipped by a scroller
- content clipped equally top and bottom, nothing scrolls →
flex-1centring, notmin-h-full - three columns on one screen, two on a bigger one → viewport breakpoint measured the window, not the pane; use container queries
- a redirect that renders an empty page and never moves the URL → do it in the router config, not inside a nested segment
- a value imported from a
"use client"module is not an array → it is a client reference; move the constant or make the consumer a client component
Fix the cause. If you also have to work around a symptom, say which is which in the commit.
Then:
- Route every affordance for the same thing through one component (one back link, one section header, one service mark). Eleven call sites edited by hand is eleven chances to drift.
- One piece of state per control. A header that reads a prop snapshot and a footer that reads live state will disagree the moment something changes.
- Never gate the input on decoration. The greeting can stream; the box appears on a timer computed from the text, not on an event that may never fire.
Step 5 — Re-measure, then look once
Run the exact evals from Step 2 again. Numbers first, then one screenshot per
changed screen. If a screenshot contradicts a number, the screenshot wins — find out
why before moving on.
Do not call a flow done from one width, one path, or one screenshot. Two of the three worst defects this skill was written from were only visible at 577px tall and only on the click-through path.
Step 6 — Report honestly
Lead with what was driven (in a browser, against real data) versus what was only compiled. Give the measurement, not the adjective:
Continue's right edge sits on the box's right edge (1027 = 1027). Meet your agent is at 532 in a 577 viewport. Strip is the same DOM node across navigation; the brand loading screen never appeared.
State what you could not reproduce and what you did not test. A user who is told "I couldn't reproduce this at 1440 or 400 — send me your window size" can help you; one told "fixed" cannot.
Write the record to {output_dir}/verify-YYYY-MM-DD-[slug].md in break-it and overhaul
modes: defects ranked, each with what you did / expected / got / proof, then the fixes
and their re-measurements.
Hard rules
- A screenshot from the user outranks your reading of the source.
- No "done" without a number from the rendered page.
- Four widths, the click-through path, a reload, and one mid-animation frame — every time.
- Search patterns.md before diagnosing from scratch; add to it when you find a new one.
- One component per recurring affordance; one state per control.
- Never commit a test file you did not run after editing it.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

