A skill that makes Claude review a new piece of work the way a brand's design director does before it ships. It turns your brand guide, or your best past work if there is no guide, into a one-page rules sheet, with every rule marked as written, observed or a default. Then it takes the ad, post, page, pack or deck apart and checks the logo, colour, type, imagery, layout, voice and claims against those rules in a pass/fix table. Every finding names the rule it breaks and the number behind it, and every fix is exact: the colour value, the size, the position, the new line with its character count.
Inside are a rules sheet template with defaults for where a guide is silent, visual checks with figures for logos, colour, type, imagery and layout, and a tested Python script that measures colours against your palette, works out WCAG 2.2 contrast and reads the fonts inside a PDF. There is a voice and copy reference, and a guide to claims that need proof under US rules, covering FTC endorsement and review rules and FDA wording for cosmetics and supplements, each with its source and the date it was checked. It never invents a rule, a measurement or proof, and it never tells you a claim is legal.
It flags claims and says when a lawyer should read the piece; it does not give legal advice, and the claims check covers the US only. It cannot open InDesign, Illustrator, Figma or Photoshop files, so send a PDF or PNG export, and it cannot judge printed colour, which needs a proof. Without code execution it judges colour and type by eye and says so. For the best result, send the whole set of pieces at once with the guide, the caption and whatever proof you hold for your claims, then run it again on the revised version before it goes live.
Claude skill · version 1.0
On-brand checker
Checking your access.
In the download · 9 files
- SKILL.mdThe method Claude follows, stage by stage: the rules sheet, the inventory, the pass/fix table, the claims register, the fixes and the sign-off.
- README.mdHow to install it in claude.ai or Claude Code, how to start, and the licence.
- references/brand-rules-sheet.mdHow to turn a brand guide, or the brand's best past work, into a one-page rules sheet, with every rule labelled written, observed or default, and the defaults for where a guide is silent.
- references/visual-checks.mdLogo, colour, type, imagery, layout and accessibility checks with figures, WCAG 2.2 thresholds and platform crop zones with sources, and a tested Python script that measures colours, contrast and PDF fonts.
- references/voice-and-copy.mdVoice attributes, tone by channel, house style, terms and inclusive language, generic-copy tests, severity for word findings and how to write a fix.
- references/us-claims.mdClaims that need proof under US advertising and labelling rules (FTC, FDA, USDA, USPTO), endorsement and review rules, pack and email requirements, each with its source and the date checked.
Install it and begin
01Install it in Claude CodeTerminal mkdir -p ~/.claude/skillsunzip ~/Downloads/on-brand-checker.zip -d ~/.claude/skills/# Start it with /on-brand-checker. If this is your first skill, type /reload-skills once.On claude.ai there is nothing to unzip: upload the zip as it is. Every step is on Install and use a Claude skill.
02Your first message to ClaudeText Use the on-brand checker skill. Check the attached [PIECE, E.G. INSTAGRAM AD] against our brand guide. It runs on [WHERE IT RUNS] in [MARKET, E.G. THE US] and goes live on [DATE]. Start with the rules sheet.Attach the piece as a PDF or PNG export and your brand guide, or ten of your best past pieces if you have no guide. Turn on code execution so Claude can measure colours, contrast and fonts.
Checklist
Before a piece ships on brand
- Every finding points at a written rule in the guide, or is labelled as a default or as taste.Without a rule, a review is an opinion, and opinions get argued with instead of fixed.
- The cover test: with the logo hidden, the piece is still recognisably yours.Every rule can pass and the piece can still look like the rest of the category. This catches it.
- Colours measured against the palette, within 2.0 Delta E 2000 on screen files (the skill's working limit when the guide sets none), with the accent kept to its stated share.In-between tints and accent creep are the commonest colour drift, and eyes miss both.
- The headline font read from the exported PDF, not judged from a screenshot.Template tools swap a missing font for a lookalike without telling anyone.
- Text contrast at 4.5:1 or more at the size it shows on a phone, where a 28 px line on a 1080 px post is small text.Contrast is judged at the size people read it, not the size it looks in the design tool.
- The product label in every image correct letter by letter at 200%.Image tools redraw labels, and a misspelt brand name on the pack undoes the whole ad.
- Every claim listed with its proof on file, and creator posts disclosed before the "more" cut.US rules expect proof before an ad runs, and a disclosure hidden after the cut does not count.
- Any waived rule written down, with who accepted it.The next reviewer can tell what was decided on purpose from what slipped through.
What to do it with
- ClaudeFree tier, paid above it
A language model that writes, edits, reads long documents and runs tools, including agents that work inside a codebase.
Bad at It is agreeable by default, so it will improve a bad idea instead of telling you it is bad. Ask it to argue against you and the quality of the answer changes.
- Claude CodeSubscription
An agent that reads a whole repository, edits files, runs the tests and explains what it changed.
Bad at It is fast enough to make a large mistake quickly. Work on a branch, read the diff, and never let it push. Watch the starting mode: on Pro, Max and Team plans it opens in auto mode, which Anthropic's own list says may read a .env file and send the key in it to the API that key belongs to. Switch it off.
Tells you whether two colours are legible together, against the published standard.
Bad at The ratio is a floor and not a design. Passing text can still be unpleasant to read.
- WhatTheFontFree
Identifies a typeface from a photograph of it.
Bad at It will confidently name a close relative rather than the actual face, and it only knows what its own shop sells.
- Canva Brand KitsFree tier, paid above it
Locks fonts, colours and logos so everybody making a post uses the right ones.
Bad at It enforces assets, not judgement. A team can stay perfectly on-palette and still produce work that is off-brand.
Why it matters
Asked whether something is on brand, a general chat gives an opinion: it feels premium, maybe make the logo bigger. Work fails on smaller, checkable things: a font swapped for a lookalike, the accent colour used as a background, white text on a mid-tone that fails contrast, a creator quote with its disclosure hidden after the cut, a "clinically proven" nobody can back. This skill makes Claude measure each one against your own rules and hand back exact fixes.
- Type
- technique
- Period
- ongoing