Three Screen Sizes
Every page is captured on two laptops and a phone
Each page is opened at 1440×900, 1280×800 and 390×844, the way a person would see it. A layout that only works on your own monitor gets caught.
Every page is captured on two laptops and a phone
Each page is opened at 1440×900, 1280×800 and 390×844, the way a person would see it. A layout that only works on your own monitor gets caught.
Flags any page that scrolls sideways
A page that is wider than the screen fails, with the number of extra pixels. Wide tables are expected to scroll inside their own box instead of dragging the whole page.
Flags text smaller than 16 px, even inside charts
Every piece of visible text is measured, including chart labels at the size they actually render. The floor is 16 px by default, and you can raise it for an older audience.
Buttons and links under 44 px get flagged
Every button, link and form field is checked for a 44 px tap area both ways. Links inside a sentence are left alone, since nobody expects those to be big.
Catches text cut short while there's room beside it
Titles that end in an ellipsis even though there is plenty of space next to them get flagged. You see the full name unless the screen truly has no room for it.
Catches words split in the middle across two lines
A name that breaks into "Wildflo" on one line and "wer" on the next gets flagged, along with the CSS fix. Long names with slashes get a clean break point after the slash instead.
Counts how many items fit on the first screen
On lists, tables and stores, it counts what you see before scrolling and how much space the header takes first. Targets like six list lines on a laptop and three on a phone turn "feels empty" into a number.
Desktop content should fill at least 70% of the width
A narrow column floating in a wide screen is flagged with how much of the width it uses. Content is expected to fill most of the screen, capped near 1200 px and centered.
Full, empty, long text and missing values each get a look
Each page is captured in every state a person can reach: full of data, empty, with very long names, with missing values, and as each kind of user. You list the pages and states in a small pages file (an example is bundled). The empty state is where layouts usually break.
Sheets, dialogs and expanders are opened and measured
Anything that only appears after a click, like a slide-up sheet, a dialog or an expanding section, is opened the way a person would and measured as its own state.
Signs in through your real form, password never shown
Pages behind a login are reached by signing in through the actual sign-in form, using a small sign-in hook you write (an example is bundled). The hook reads the password from an untracked env file and never prints it, and any secret on screen is blanked before the screenshot.
Tests with lifelike data, then restores what was there
Before capturing, the feature is filled with data shaped like real use: realistic counts, long names, blanks. The starting state is noted first and put back afterward.
Checks every page your change touches, not just one
Point it at a few pages, or give it none and it checks every page the current branch changes. You don't find the next broken page yourself after it said the first one was fine.
One line per page and size, with what failed
A short table lists every page, state and screen size with its failures and counts. You see at a glance which page breaks where.
A first-screen and a full-page picture of each
Every page, state and size gets two screenshots, what you see before scrolling and the whole page, named so they line up. They become the evidence for every finding.
Claude opens every screenshot itself
Numbers miss things like a label sitting on a bar, two items both called "Other" or a missing way back. Every screenshot is looked at directly, plus the in-between tablet widths where layouts switch.
Problems listed by how much they hurt
The audit has a pass/fail table per page, then the problems ranked by how much they get in a person's way. Each one names the page, size, screenshot, number and a proposed fix.
Density targets for data pages, judgment for articles
Lists, tables, stores and dashboards get the item-count targets. Articles and landing pages keep the text, tap, scroll and cut-off checks, with density judged by eye.
Each problem becomes a design review with a real screenshot
Every problem, or group of problems with one fix, becomes an item in a design review. The Before is the real screenshot, or the one you sent if you sent one.
Each proposed fix visibly meets its target
Mockups use your page's own fonts, colors and data, and are drawn to show the target being met: a laptop frame that really holds eight lines, a phone frame with 16 px text and 44 px buttons.
Every item states the failing number and the target
Each review item's reason gives the measured number next to the target it missed. You decide on facts, not on how a page feels.
Choices that aren't layout fixes become questions
Anything that is really your decision, like whether to hide sold-out items, is asked as a question rather than quietly decided as part of a layout fix.
Fixes are re-measured and shown side by side
After the approved fixes are built, the same pages are captured again with matching file names. You get Before and After screenshots side by side with the key numbers, before anything is merged.
Numbers worth keeping get pinned in a test
A number that matters, like how many rows fit or how many database queries a page makes, can be pinned in an automated test so a later change can't quietly undo the fix.
A private link, a local file, or your Wi-Fi
Your project picks where review pages go: a private claude.ai link, an HTML file that opens in your browser, or that file served on your home Wi-Fi so a phone can open it. Nothing is ever uploaded or tunneled in the Wi-Fi mode, and the server stops when the review ends.