How to extract your brand colors and voice for any AI tool

The first time most operators try a generative AI tool for branded social content, they get coffee. Specifically: a generic latte on a wooden table, pleasant beige tones, no specific identity. The output is technically fine and completely useless. The fix is a brand profile, a structured paste-in block of colors, voice, photo style, props, and forbidden patterns. Once that profile exists, the same tool produces work that actually looks like the brand. This post is the playbook for extracting that profile from a real business and feeding it into any AI tool that will accept structured input.

The five inputs that matter

Five categories carry almost all the weight. Anything else is decoration.

Colors. Two to four hex values, each with a role label: primary, secondary, accent, neutral. Without role labels, the model averages the palette and you get a muddy version of every color at once. The roles are what tell the model which color belongs on the cup, which goes on the linen, which is the small surprising accent in the corner.

Voice. Three adjectives and three anti-patterns. "Confident, unhurried, specific." "Not hype-y, not casual, not corporate." The anti-patterns matter as much as the adjectives because they prevent the model from defaulting to its trained-in average voice (slightly excited, slightly generic, faintly LinkedIn-shaped).

Photo style. Four sub-fields: lighting, composition, depth of field, mood. "North-facing window light, scene-first composition, shallow depth of field, Kinfolk magazine mood." This block is what stops the model from defaulting to the AI photography aesthetic (overhead flat lay, oversaturated, perfect symmetry, empty surface).

Props. Three to seven specific objects that visually carry the brand. A navy ceramic cup. An oat linen apron. A brass espresso group head. A single terracotta planter. These are the small repeating elements that make the feed feel like a feed, rather than nine unrelated photos.

Forbidden patterns. A short list of "things that have ever made me cringe on this brand." Generic stock photography. Posed eye-contact smiles. Chrome espresso machines. Quote graphics on solid color backgrounds. Whatever the equivalent is for your brand, write it down. The model will not infer it on its own.

The same prompt, with and without a profile

Here is what the difference looks like for a single lifestyle shot. The left side is a naked prompt typed into any AI image tool. The right side is the same prompt, but the Bluebird Coffee brand profile has been appended to the prompt before it hit the model.

Prompt

Result

The naked prompt is honest about what it asks for and the model gives back exactly what it asks for: a generic couple, a generic cafe, a generic latte. The branded version asks the same thing but constrains the model's choices on every other axis. Same subject, recognizably specific brand.

This pattern works in any AI tool that accepts a freeform text prompt. The brand profile is not a feature of one product. It's a discipline you apply to every prompt you ever send.

Brand uploads: when an image is your brand

Some brand assets cannot be described in text. A logo is the obvious one. The rose-pink sun-emblem from IL Gelato Hawaii (a real ice cream shop on the North Shore of Oahu) cannot be reconstructed from "rose-pink sun with rays in a hand-drawn style." It needs to be uploaded as a reference image alongside the text prompt. Most modern AI tools accept inline image inputs. Sevenposts wires that path automatically.

The pattern is the same as the text profile, just with images: feed the model the asset, then tell the model what to do with it.

Logo

Applied

Three asset types are worth uploading rather than describing:

1. Logos and wordmarks. The model will never reproduce them correctly from a text description. Always upload.
2. Signature products. The thing on your bestseller page that customers recognize at a glance. The exact bag, bottle, or device. Describing it costs precision.
3. Founder portrait or hero image. If your feed has a recurring face, upload one good photo of that face. The model will use it for likeness without copying it pixel-for-pixel.

Other assets (storefront, signage, packaging) are usually faster to describe than to upload. Use uploads where the asset is unique enough that no description would land it.

The structured profile template

Below is the actual paste-in template. Save it as a snippet, fill it in once per brand, and append it to every prompt you send. The order matters: colors first because they carry the most visual weight, forbidden patterns last because they're the strongest constraint and the model attends to the most recent context most heavily.

## colors
- primary: #1f3a5f (deep navy, used on ceramic cups and packaging)
- secondary: #e9d8a6 (oat linen, used on aprons and napkins)
- accent: #c8553d (terracotta, used sparingly, one element per scene)
- neutral: #fafafa (warm off-white, surfaces and negative space)

## voice
adjectives: confident, unhurried, specific
anti-patterns: hype-y, casual, corporate

## photo style
lighting: north-facing window light, no flash, no harsh shadows
composition: scene-first, subject second, product third
depth of field: shallow, background softly blurred
mood: Kinfolk magazine, editorial restraint

## props
- navy ceramic cup
- oat linen apron
- brass espresso group head
- reclaimed wood counter
- terracotta planter (one per scene, sparingly)

## forbidden
- chrome espresso machines
- posed eye-contact smiles
- generic stock-photo coffee
- quote graphics on solid backgrounds
- overhead flat-lay symmetry

That's the whole thing. Roughly 25 lines. It compresses a year of brand-discipline decisions into a block the model can read in one pass.

How to extract each input from your existing brand

Most operators have already done the work. They just have it scattered across Figma files, Instagram drafts, and a moodboard nobody opens. Walk through it once, in order, and write each input down.

Colors. Open your logo file in Figma or any vector tool. Use the eyedropper on every distinct color. Write the hex code and a one-phrase description of where it appears on the brand. If your logo only has two colors, pull the other two from your three most-saved Instagram posts.

Voice. Open your Instagram and scroll to the three captions you're proudest of. Copy them into a text file. Read them out loud. The shared adjectives are your voice. For the anti-patterns, read three captions you regret. The shared mistakes are what you're trying to avoid.

Photo style. Open your feed grid view. Pick the nine images that look most "you." Open them in a folder side-by-side. Name what's consistent: the light direction, the framing distance, the depth of field, the overall mood. That's your photo style.

Props. Scroll your last 30 posts. Note every object that appears in three or more frames. Those are your props. Anything appearing in exactly one post is a one-off and doesn't belong in the profile.

Forbidden patterns. This one needs a notebook moment. Sit down for ten minutes and write everything that has ever made you wince when you saw it on the brand. Other people's mistakes count too. The list will be specific and slightly petty. That's the right shape.

The whole walkthrough takes about an hour for a brand that has been running for at least six months. For a newer brand, give it two hours and accept that the first version will be partial. Profiles iterate. The first draft does not need to be the final one. It needs to exist.

What to do this week

Pull out a notebook and walk the five inputs in order. Don't aim for a polished version. The first draft of a brand profile is always rough, and the only way to find the gaps is to paste it into a prompt and see what comes back. Iterate from there.

For more on the broader question of what "on-brand" even means as a discipline, the four pillars framework is the next read. It covers colors, voice, photography style, and cadence as the four axes a brand has to be coherent across.

For the failure mode that motivated all of this, why generic AI image generators fail for product brands explains why a model with no profile defaults to the same averaged aesthetic regardless of what you ask for, and what that aesthetic actually looks like.

If the brand profile lands and you want to see what the full workflow looks like with one in place, the 60-second social post walks through one end-to-end run from prompt to scheduled post.

A working brand profile compounds. Every prompt gets cheaper to write because the work was front-loaded once. The model gets more useful, not less, as the profile gets sharper. For a deeper take on the same idea applied to enterprise brands, the public IBM Brand Center is the published reference most large brands have already done. Yours doesn't need to be that thorough. It needs to fit in 25 lines and live in a snippet you can paste in three seconds.