How to keep AI-generated content on brand

Brand guidelines were written for people

A new designer reads the brand PDF once, then learns the rest over months: from feedback, from reviews, from the person at the next desk. An agent gets none of that. Each run starts with only what’s in front of it, so drift comes from what it was given:

No brand at all

A new chat, teammate or agent starts from nothing and falls back on the most common choices. That’s why unbriefed output tends to sound and look like everyone else’s.

Adjectives

“Friendly but professional” describes most brands. A model can’t check its work against an adjective, and neither can you.

The wrong form

A palette described as “warm and earthy” gets approximated, and a logo described in words gets redrawn. Many brand prompts cover tone and stop there, leaving the visual identity to guesswork.

Copies

Every pasted prompt, saved instruction and tool setting is a copy of your brand. Copies fork the day the brand changes, and nobody updates all of them.

Brand works with agents when it’s operable: specific enough to apply, delivered in the right form, checkable and current. The rest of this article is how to get there.

Turn adjectives into rules

Good guidance is specific enough that someone else could check the result against it. Replace adjectives with rules, and give the reason when it isn’t obvious. Models apply a rule more sensibly when they know what it’s for.

Add a few examples of finished work you’d approve, and one or two you wouldn’t, with the reason. Keep the stable brand separate from campaign details, which change every few weeks.

Give each output the brand in its own form

A full brand book pasted into every request looks thorough, but it leaves the model to work out which of 60 pages matter for one post. Each kind of work needs different parts of the brand, in a different form.

Copy

Emails, posts, product text

Voice rules, terminology, product facts, the claims you can and can’t make, and a few approved examples.

Layouts

Slides, documents, web pages

Exact values and files: hex codes, font files, logo files and any templates. A tool that can read #F2C94C shouldn’t be guessing at “warm yellow”.

Visuals

Images, video, SVG

Show the brand rather than describe it. Attach approved images, the logo file and type specimens; use the prompt for what to make and the exact words.

For visuals, references do most of the work. OpenAI’s image guide recommends naming what each reference is for, such as subject, style or background, and what a style reference should carry over: “its palette, texture, or visual medium”. Google’s Gemini image models accept several reference images in one request.

Plan for two limits: models can misspell a brand name or redraw a logo slightly wrong, so spell unusual names letter by letter, as OpenAI suggests, and place the approved logo file on the finished image instead of asking for a redraw.

Picking the right parts by hand for every task doesn’t scale. That’s the job of a Rainbrand Brief: your agent describes the task and gets back the guidance and assets it needs, then makes the work with your own tools or through the AI Gateway.

Try it with your own brand in place of Acme:

Request a Brief for Acme: a 1080 × 1350 Instagram post launching Atlas Pro, with the headline “Atlas Pro is here.” Make the image with your image model, using the Brief’s reference images for style, and place the logo from its approved file. Then request a Review of the image.

Check against the brand, not a feeling

“Is this on brand?” invites a general answer, whether you ask a colleague or a model. Check against specific rules instead, and keep “not sure” separate from “passes”.

A Rainbrand Review runs checks like the visual ones against the brand itself and shows evidence for its findings where available, so your agent can revise the work and check it again.

Keep one brand that every agent calls

Copies are where drift starts. The alternative is one maintained version that every agent and product reads from, kept alongside the tools and sources you already use.

In Rainbrand, that version is the brand itself. Agents reach it through MCP, the open standard for connecting AI applications to outside systems, and products reach it through the API, so a Brief requested from Claude, a script or your own product draws on the same brand. When the brand changes, you describe the change, review the update Rainbrand prepares and apply it, and every new Brief draws on the updated brand.

Retire assets the same way. When a logo, product or campaign is retired, take its files out of the Brand Library, or an agent will keep finding them.

Many brands, one layer

Agencies, studios and products that serve many brands have one more failure: the right guidance attached to the wrong brand. Keep each brand’s context separate, name the brand in every request instead of relying on a default, and check the brand itself as part of review. A sub-brand with its own voice is a separate brand for this purpose.

In Rainbrand, each brand keeps its own context, shared with your team where you need it, and agents name the brand they’re working on in every request.

Start with one brand

If brand work runs through several agents, people or brands, start with one brand and your next real task.

Connect your agents

Add Rainbrand to the agents you already use.

Add a brand

From whatever defines it: your website, brand guide, logos and images.

Request a Brief

For that task, then make the work with your own tools or the AI Gateway.

Request a Review

Check the result against the brand before it ships.

Results still depend on your sources and the models you use, so keep the check in the loop.