- Home
- Skills
- Frontend & Web Apps
- Proofed UI
More screenshots
Works with the AI Agents you already use
Proofed UI
A design and redesign skill that checks the real page in a browser and scores it before and after.
$3
Proofed UI
Example session with this skill installed
Redesign this landing page so it looks and works better. Keep the content honest and every feature working, and show me the before and after.
- Read your context and instructions
- Compiled the proofed ui
- Generated the app files
Fieldnote landing page: redesign report
Example from a real run of Proofed UI's redesign mode on a fictional test page (Fieldnote is a made-up product). Local file paths are shortened, and the diagnosis file it mentions is not included. Nothing else is changed.
Date: 8 November 2026
Page: index.html of the test page, opened directly as a file
Files changed: index.html, styles.css, app.js, IMPLEMENTATION.md
In short
Proof score: 45/100 before, 100/100 after. Both runs used the same renderer, the same browser, the same page and the same three screen sizes.
- All 39 problems the rendered checks found are gone: 25 text-contrast failures, 8 small tap targets on phones, a skipped heading level and a repeated "small uppercase label" pattern. The one signal in the code, default fonts only, is gone too.
- Every behaviour in IMPLEMENTATION.md still works. A test page ran 44 of 44 checks on the original page and 48 of 48 on the new one; the 4 extra checks cover the new day sign.
- One small visual issue remains. On phones the demo's "The day ahead" heading wraps onto two lines next to the Reset button (see "What's left").
The Proof score counts what a fixed set of checks found. It does not measure beauty, sales or full accessibility compliance. The screenshot review and the manual checks listed at the end cover more of that.
What was wrong with the original page
Measured by the checks
Grey text that was too faint. The secondary grey (#677369) passed only on the plain cream background. On every tinted panel it fell below the 4.5:1 minimum: the "How it works" strip (3.62:1 and 4.35:1), the demo sidebar (2.83:1 and 4.29:1) and the selected visit card (4.41:1). That made 25 high-severity findings across the three sizes.
2.
Controls too small to tap on a phone. The "Prepare a demo enquiry" link was 23px high and "Back to top" 16px. The Reset button, both demo dropdowns, the name field, the crew-size dropdown and the footer logo link were 32–39px high, under the 44px guideline.
3.
A skipped heading level. The page went from its main heading straight to three third-level headings, with no section heading in between.
4.
The same small spaced capitals over most headings ("01 / PLAN", "TAKE IT FOR A SPIN", "LET'S TALK ABOUT YOUR DAY", "VISIT DETAILS" and more). It's a stock template pattern, and most of the labels repeated their heading.
5. Default fonts only (Arial and Georgia).
Seen in the screenshots and the code
Tiny text everywhere. Real content was set at 7–11px: visit times, addresses and crew names at 9px, status badges at 8px, the "Interactive demo" pill at 7px on phones.
7.
The hero art cut off its own label. The vertical "ROOM TO GROW." read "OOM TO GROW" at every size. The illustration said nothing about scheduling, and on phones it pushed the working demo about 1,250px down the page.
8.
A main button that went to the wrong place. "Explore a demo" in the header opened the enquiry form, not the demo, and on phones it was the only link in the header. Links that stay on the page carried the ↗ "opens elsewhere" arrow.
9.
A cramped demo. It had a decorative sidebar with one menu item and 8px text. On phones the three visit cards were squeezed into three 100px columns.
10.
Vague headings. "Good days start with a clear plan." and "Make space for the work outside." don't say what the product is or what the form does. On phones the forced line break left "start" alone on a line of the main heading.
Nothing on the page was an unverified claim. It had no statistics, customers, testimonials, logos or prices, and it already said clearly that the product is fictional. That's unchanged.
What changed, and why
I kept your palette (deep green, cream, peach and terracotta), the name and the asterisk mark, and changed the type, the layout and the hero. Before editing, I compared three directions. The one chosen is a
day sheet with park-signage type: show the sample day itself, the product's real object, rather than a decorative picture. I rejected a "field notebook" look, because handwriting fonts lose legibility outdoors and it puts the note ahead of the plan. I also rejected a "site plan" drawing, because it speaks to garden designers and hints at map features the product doesn't have. The full diagnosis and plan are in diagnosis-and-plan.md next to this report.
1. Colour, type and sizes (styles.css)
- The secondary grey is now
#535f55, at least 5.1:1 on every light surface on the page. Every text colour was checked against the surface it sits on. - Headings, times and labels use
Bahnschrift, a DIN-style face that ships with Windows 10 and 11. It looks like trail signs and work orders, which suits crews working outside, and its figures make times like 08:00 easy to scan. Reading text uses the system UI face (Segoe UI on Windows). Nothing is downloaded. Other systems fall back to DIN Alternate (macOS) or their own UI font.
- Body text is now 16px, and nothing on the page is smaller than 13px.
- All buttons and form fields are 44–48px high. The focus outline is a clear 3px terracotta ring.
2. Header and hero
- The header button now reads
"Try the sample day" and goes to the demo. "How it works" and "Demo enquiry" are plain links beside it. On phones the header keeps the main button, and the hero's two buttons cover the demo and the enquiry.
- Removed the ↗ arrows from links and buttons that stay on the page.
- New heading:
"A clear plan for every crew's day." The intro now names the product and the audience: "Fieldnote is a day planner for small landscaping crews. Every visit gets a time, a crew, a status and a note for whoever goes next."
- The decorative garden and its clipped label are replaced by the
day sign. It's an arch-topped green panel that keeps the arch shape of the old artwork, and it lists the three sample visits with their time, crew and status. It's built from the demo's own data, so it updates when a visit is saved in the demo below. On screens 600px and narrower it is hidden, because the demo's own list comes right after the hero there.
3. How it works
- The section now has a proper heading, which fixes the skipped level, and an ordered list of three steps marked by large numbers instead of small capitals.
- The step titles are kept. Their descriptions now match what the demo actually does: visits in time order with address and work type; a crew and a status per visit; a note that stays with the visit.
- On phones the steps stack in one column instead of three narrow ones.
4. The demo
- The decorative sidebar is gone. Its "fictional data" note moved into the line under the demo: "A working mock schedule with fictional data."
- The visits sit on a
time rail, with start and end times on the left and a marker per visit: an open ring for planned, amber for in progress, green for completed. The status is also written out next to the crew name.
- Cards are full width on phones and tablets. The toolbar keeps the day, the "0 of 3 complete" count and a Reset button you can comfortably tap.
- Smaller fixes: the visit list is labelled as a group for screen readers. Each card's spoken name now includes its time. The note field and the enquiry text box now show their character limits.
5. Enquiry and footer
- The heading now says what the form does:
"Prepare a demo enquiry". The note "This is a local prototype… never sent" stays word for word.
- The footer text is readable, and both footer links are full-size tap targets.
What stayed the same
- All four sections and their addresses:
#home,#how,#scheduleand#enquiry. - Every id and behaviour in IMPLEMENTATION.md.
- The sample data in app.js: three visits, two crews, the same times, addresses and notes.
- The fictional-product notices, and the enquiry that is prepared in the page and never sent.
- No network requests, storage, packages, web fonts or images.
- Kept word for word: the top banner, "A little less 'who's on this?'" and its instruction line, the three step titles, the form labels, the validation and feedback messages, the summary text and the footer line.
- IMPLEMENTATION.md now describes the new day sign (
#day-board,#day-board-list,#day-board-count), the cards'data-status, the hint ids and the font choice.
Before and after
| Measure | Before | After |
|---|---|---|
| Approach and partial | renderer, partial: no | renderer, partial: no |
| Pages × sizes scored | 1 page × 3 sizes (375×812, 768×1024, 1440×900) | 1 page × 3 sizes (same) |
| Combined Proof score (score.json) | 45/100 | 100/100 |
| Lowest page and its score | index.html: 46 | index.html: 100 |
| Rendered high / medium (finding entries, all sizes) | 25 high / 5 medium (plus 9 low) | 0 / 0 (and 0 low) |
| Tells in the code (distinct ids) | 1 (S-FONT-DEFAULT-ONLY) | 0 |
| Console errors | none | none |
| Failed or redirected pages | none | none |
Checks that could not run (errors) | none | none |
Before, by rule: R-CONTRAST 25 high; R-TAP-TARGET 2 medium and 6 low, phone only; R-TELL-EYEBROW 3 medium; R-HEADING-ORDER 3 low.
Measurement conditions were the same in both runs: Chrome 157.0.8103.38 on Windows 10, Node 24.20.0, the same single page opened as a file, the same three sizes in the same order, with nothing loading from the network. Only the page changed, and coverage is identical, so the comparison is like for like.
Screenshots
The before and after files have the same names. Fold screenshots (1-<size>-fold.png) sit next to each full-page one.
| Page and size | Before | After |
|---|---|---|
index.html, mobile | …\proofed-runs\proofed-ui-before\1-mobile-full.png | …\proofed-runs\proofed-ui-after-3\1-mobile-full.png |
index.html, tablet | …\proofed-runs\proofed-ui-before\1-tablet-full.png | …\proofed-runs\proofed-ui-after-3\1-tablet-full.png |
index.html, desktop | …\proofed-runs\proofed-ui-before\1-desktop-full.png | …\proofed-runs\proofed-ui-after-3\1-desktop-full.png |
… stands for a temporary folder outside the project. These are temporary folders: copy anything you want to keep.
How it was checked
- Baseline
proofed-ui-before: run before any edit. - Repair rounds, each rendered into a new folder:
proofed-ui-after-1, the full redesign: 100/100. Visual review found inconsistent status badges and cramped cards on phones and tablets, plus a one-word caption line.proofed-ui-after-2: 100/100. Visual review found "Reset demo" wrapping onto two lines on phones.proofed-ui-after-3, the final state: 100/100. Each folder holdsrendered.json,score.jsonand six screenshots. I looked at every screenshot of every run.
Behaviour test page. A copy of index.html in the scratchpad loads the page's real app.js and styles.css plus a test script (…\scratchpad\tools\behaviour-test.js). It was rendered with the same renderer; a failure would show up as a console error and in the page title. It covers:
- the three in-page links;
- selecting a visit and
aria-pressed; - loading saved fields, the crew and status options, the 400-character note limit, and discarding unsaved edits when switching visits;
- saving, which updates the badge, the count and the announcement;
- Reset;
- the required enquiry fields, rejecting a whitespace-only enquiry, and clearing that error on typing;
- the summary inserted as text, never HTML ("<b>Alex</b>" stays literal), and "Name not provided";
- the copy feedback;
- the day sign following saves and resets.
Results: 44/44 on the original page (behaviour-before) and 48/48 after every round (behaviour-round1 to behaviour-round3; round 3 at all three sizes). This test page is not part of the Proof score.
What's left
Unresolved
Phones: "The day ahead" wraps onto two lines beside the Reset button in the demo toolbar (proofed-ui-after-3\1-mobile-full.png). It's cosmetic, and nothing overlaps or gets cut off. I stopped because the skill allows three repair rounds. The likely fix is a smaller toolbar heading and a narrower gap on phones (in the max-width: 760px block of styles.css, .app-toolbar h3 from 1.5rem to 1.25rem and the column gap from 24px to 12px), followed by one more check.
Desktop: empty space under the visit list. The edit form is taller than the three sample visits, so the list column has some empty space below the cards. I left it, because a real day would usually hold more visits.
Not verified by these checks
Other browsers and systems. Only Chrome on Windows was rendered. On macOS the headings fall back to DIN Alternate, and elsewhere to the system UI face; I didn't see those fallbacks, or the page in Firefox or Safari.
A successful copy. In the headless browser the clipboard was refused, so the page showed its documented manual-copy message. The "Copied. Nothing has been sent." path needs a quick try in a normal browser window.
File sizes. Font and image sizes can't be measured on a page opened as a file. The page loads no font or image files at all, so this is low risk.
Manual checks worth doing
- Tab through the page with a keyboard. Check the focus order and that the focus ring shows on every control, including the visit cards.
- Listen to the visit cards and the day sign with a screen reader.
- Zoom the text to 200%, and narrow the window to 320px.
- Try the page with reduced motion turned on.
- Look at the hover states.
Facts you need to supply: none. No placeholders were added, and the page makes no claims that would need a source.
Optional next step
If you'd like later pages to match this one, I can save the colours, type and spacing as a DESIGN.md plus a small tokens file in the page folder (a "brand lock"). That needs your OK first.
index.html
ZIP · project files
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
Design changes you can measure
AI coding agents can build a page in minutes, and the pages tend to share the same problems:
- grey text too faint to read
- buttons too small to tap on a phone
- content wider than the screen
- the same template look as every other AI page: small uppercase labels over every heading, purple gradients, "Get started / Learn more", made-up stats
Proofed UI makes your agent check the real page in a real browser at phone, tablet and desktop size. The agent fixes what the checks find, then shows you a before-and-after Proof score with screenshots.
Two modes
- Redesign your own site. Your agent measures the page first, explains in plain words what's broken or generic, and shows you its plan. It edits your code only after you agree, one area at a time. Then it measures again under the same conditions.
- Design a new page. Your agent researches your subject and writes three different design directions. It picks one and tells you why, builds it, then checks it the same way.
What it checks
- The rendered page, in your installed Chrome or Edge or your agent's own browser, at 375×812, 768×1024 and 1440×900:
- Readability: text contrast, clipped text, content wider than the screen, and links or buttons that overlap.
- Phones: tap targets that are too small, and a missing viewport tag.
- Structure: heading order, a missing h1 or page language, and images without a description or a size.
- Keyboard: focus styles.
- Weight: layout shift, heavy fonts and images, and too many font families.
- Template tells: repeated eyebrow labels, gradient heroes, identical card grids, emoji icons and everything centred.
- Your code, for 17 template tells:
- lorem ipsum;
- made-up social proof and stat strips;
- stock phrases ("unlock the power", "seamless");
- purple gradients and glow blobs;
- starter leftovers and builder branding;
- default fonts only;
- three icon cards and generic "Get started" buttons.
What you get
- A Proof score from 0 to 100, computed by fixed rules from what the checks find, with screenshots at all three sizes.
- A plain-English report: what was wrong, what changed and why, what's left, and what the checks can't judge.
- Brand lock, if you want it: your colours, fonts and spacing saved as DESIGN.md plus a tokens file, so later pages match.
How it was tested
We tested it on fictional pages made by separate AI agents. Every page was scored at all three sizes in Chrome on Windows.
- Redesigns of two AI-made landing pages: they went from 45 and 46 to 100 and 100 with Proofed UI. The same AI model, given a plain "redesign this" request, reached 85 and 85. It fixed the contrast and sizing too, but kept the template habits.
- New pages from three briefs (a bakery, a B2B tool, a tattoo artist): a mean of 97.5 with Proofed UI, against 63.7 without. One page built without it also scored 96.
The score counts checkable problems. It doesn't measure taste or originality: you still choose the look, and the report gives you screenshots to judge it.
Safe by design
- Nothing to install. Two single-file Node.js scripts with no dependencies. They make no internet requests.
- Screenshots go to a temp folder outside your project. The browser runs with a temporary profile that is deleted afterwards. A force-stopped check can leave that temp folder behind, and it is safe to delete.
- No silent changes. A redesign edits only after you OK the plan, and brand lock is a separate yes.
- Your sites only. It declines to redesign or clone a site you don't own.
Works with
- Pages: any page your dev server serves, or a plain HTML file. For framework apps (React, Next.js, Vue, Svelte, Astro), the agent checks the dev server's address.
- Code scan: reads HTML, CSS, SCSS, Less, Sass, JS, TS, JSX, TSX, Vue, Svelte, Astro and MDX files.
- Agents: built for agents that can run Node.js and shell commands.
- Tested: in Claude Code, on Windows, with Chrome and Edge, on plain HTML, CSS and JavaScript test pages.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Frontend & Web Apps.
- For web pages and web apps, not native mobile apps or Figma files.
- Login-only screens: the renderer never signs in. Your agent checks those pages in its own browser while you sign in yourself.
- The phone-size check emulates a phone's screen and touch, but not a phone browser's identity. A server that sends different markup to phones may send desktop markup to that check.
- Redirects inside a single-page app's hash route (#/login) are not reported. Compare the screenshots with the page you meant to check.
- Tested on Windows with Chrome and Edge, on Node.js 24. macOS and Linux are untested. On those systems, run the built-in self-test first; if it fails, your agent's own browser can run the checks.
- The score counts checkable problems. It does not measure beauty, originality, sales or full accessibility compliance, and it can flag a design choice you made on purpose.
- The code checks use patterns. They skip what they can't judge reliably, but they can still miss or misread unusual code.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 3 days ago
Windows only
- Software
Node.js 18+
- Optional
- Chrome, Edge, BraveorChromium
Tested on Windows with Chrome and Edge. Without a browser, only your code is checked and the score is marked partial.
Permissions
File Scopes
render-check.cjs starts your installed Chrome or Edge with no window and a temporary profile, loads your pages at three sizes and saves screenshots to a temp folder outside your project. tells-scan.cjs only reads your code. Your agent edits files only for a redesign plan you approved, a new page you asked for, or a brand lock you said yes to. The scripts make no internet requests; the pages being checked load their own fonts, images and scripts, as in any browser.