What is the ui-verification skill?
The ui-verification skill is a SKILL.md from the mblode/agent-skills repo. It owns the browser session: your coding agent runs probes for focus order, target size, layout shift, overflow, failure states and axe violations, then reports reproduced, not-reproduced or unknown findings with evidence.
ui-verification is a skill in the mblode/agent-skills repo. Where its sibling UI skills reason about source code, this one loads the page in a browser and measures what a user would see, such as a 40 pixel control or a retry button wired to nothing.
Its output is findings tied to rule ids with reproducible evidence, plus a clearing re-run after a fix. It also has an exploratory mode that uses a finished feature like a real tester.
What changes in your agent's output
- Nine probes: axe scan, target size, focus walk, layout shift, viewport stress, failure injection, theme and locale matrix, console and network, and web vitals
- Each finding is marked reproduced, not-reproduced or unknown, never silently passed
- Measurements carry units and conditions, for example 44 by 44 pixels at 360 pixels width
- Skipped probes are listed with the reason, so missing coverage is visible
- After a fix, the identical probe runs again and the result is recorded
The skill assigns no severity tier or ship verdict. It hands evidence back to the repo's ui-design audit skill for tiering. If you only need source level checks, fixing-accessibility works without a browser.
Who it is for
Best for
- Confirming or killing suspected UI findings with a measurement
- Re-checking a fix by repeating the same route, viewport and theme
- Trying to break a finished feature before it ships
Not for
- Reading source to find defects, which the repo assigns to its ui-design audit mode
- Writing a durable test suite or maintaining pixel baselines
- Field performance data. Lighthouse style numbers are lab measurements
Use it with UiChemy
Run it against a WordPress page you built with UiChemy once it is live or on staging. Probes for overflow at 320 pixels, target size and focus order will show what the Elementor or Gutenberg output actually does.
How to install UI Verification Skill: Browser Probes for UI Bugs
git clone https://github.com/mblode/agent-skills.gitThe SKILL.md says it needs access to the target app and browser automation, and its recipes use the Playwright page API.Clone the repo
Run the command above. The skill is skills/ui-verification, with SKILL.md, a probes folder, a references folder and evals.
Copy the folder
Copy the whole folder, not only SKILL.md, into your agent's skills directory, since the probes and references load from it. For Claude Code use ~/.claude/skills/.
Point it at an app
Start your app, then ask the agent to verify a route in the browser or to reproduce a UI finding.
Check that it loaded
Ask your agent to verify one route in the browser. It should set up a session, run axe, console and viewport probes, and list any probe it skipped with the reason.
Key features
- Establish the sessionPick a driver, build mode, base URL, auth and routes first. If the app will not boot, the run stops and reports no findings, not a clean result.
- Select the probe setTake rule ids handed over from an audit and run only the mapped probes, or run axe, console and viewport probes on every route when given routes only.
- Axe scanRuns axe-core per route and theme, including computed contrast, accessible names and landmarks.
- Target size and focus walkMeasures bounding boxes and effective hit areas, then runs a scripted Tab walk that checks focus rings, dialog traps and focus restoration.
- Layout shift and viewport stressRecords attributed layout shifts with the data response held open, and checks horizontal overflow and clipped text from 320 pixels up and with tripled strings.
- Failure injectionMakes the data request return 500, an empty list, malformed data or nothing, then checks the error, empty and loading states.
- Theme and locale matrixCaptures one route across viewport, theme, pseudo-locale and direction. The skill warns that emulating color scheme does nothing for class based themes.
- Console, network and web vitalsCollects console errors, failed requests and hydration warnings, and measures LCP with its element, CLS and INP on a scripted interaction.
- Three outcomesFindings end as reproduced, not-reproduced or unknown. A not-reproduced result counts as a real deliverable because it removes false positives.
- Clearing re-run and exploratory passAfter a fix the same probe re-runs and is stored as clearedBy. A separate reference covers a fresh context pass that tries to break the feature.
Frequently asked questions
What is the ui-verification skill?
It is a SKILL.md from the mblode/agent-skills repo that runs scoped browser probes on a running app. It measures focus, hit targets, overflow, themes, request failures and performance, and returns findings with reproducible evidence linked to UI rule ids, plus an exploratory testing pass.
How do I install the ui-verification skill?
Clone mblode/agent-skills and copy the skills/ui-verification folder into your agent's skills directory, for example ~/.claude/skills/. Copy the entire folder so the probes and references stay with SKILL.md. The README also shows a cross agent installer command for the repo.
What does ui-verification need to run?
The SKILL.md says it needs access to the target app and browser automation, and its bundled JavaScript recipes use the Playwright page API. Some probes need request interception, so a driver without it cannot run them and must report them as skipped.
What is the difference between this and a UI audit?
A static audit reads source and reports what the code will probably do. This skill loads the page and measures it, so a finding arrives with evidence. It does not tier findings or give a ship verdict, which the repo leaves to its ui-design audit.
Can ui-verification replace my Playwright tests?
No. The SKILL.md lists authoring a durable test suite and pixel baseline approval as out of scope. It is for one off probes and clearing re-runs. Keep lasting regression checks in your own tests.
Should I run performance probes in dev mode?
The skill advises against it. Running layout shift or performance probes against a dev server measures on demand compilation, so it recommends a production build for those. Failure injection and focus probes are fine in dev.
Is the ui-verification skill free?
Yes. The mblode/agent-skills repo is released under the MIT license. The skill is a folder of markdown and JSON files, so you only need the browser automation tooling that your agent already uses.
Which skills pair well with ui-verification?
Use fixing-accessibility or accessibility-testing to find issues in code, then verify them in the browser here. web-quality-audit gives a broader page review whose findings can tell you which routes to probe first.


