How the Brand Brain Keeps AI Output On-Brand

The Brand Brain is a structured record — audience, voice, products, banned words, CTA style — stored against each brand. SocialPower renders it into a compact block and prepends it to every AI call, so the model reads your rules before it writes the first word. It is not chat history. Nothing is remembered by accident, and nothing decays.

Most people try to keep AI on-brand by explaining themselves in a chat window. That works for one caption. It fails at scale for three reasons: the instructions get trimmed out of the context window, they only exist in your session, and they are not there at 6am when a scheduled job generates tomorrow's post.

A stored record fixes all three. It lives in the database next to the brand, it is read fresh on every call, and it applies identically whether the generation was triggered by you clicking a button in AI Studio or by the worker in autonomous mode. Your editor gets the same voice you do. So does the post that goes out on a Sunday.

It is also editable in one place. When you decide the word "premium" is banned, you add it once and every future caption across every platform obeys. You do not re-brief anybody.

The fields fall into four groups, and they act on different parts of the output. Knowing which is which stops you filling in the wrong box and wondering why nothing improved.

Strategy fields change the angle, not the wording

Industry, target audience, mission, goals, products, services and competitors decide what the post is about. If your audience field says "cafe owners in Pune buying 5kg bags", the research step proposes wholesale pricing and delivery-schedule angles. If it says "first-time home brewers who own an AeroPress", it proposes brew ratios and grind-size explainers. Same brand, same voice, completely different content calendar.

The products field is the one people skip and shouldn't. Fill it with real SKUs and real prices and the model quotes them. Leave it empty and you get "our range of premium blends" forever.

Voice fields change the sentence

Brand voice, tone, personality and writing style shape how it sounds. Writing style is the highest-leverage field in the whole record, because it is where you put the mechanical constraints: reading level, maximum sentence length, adjective limits, whether to lead with the fact or the feeling. "Sentences under 18 words, class 8 reading level, lead with the concrete fact" is an instruction a model can follow and you can check. "Professional yet friendly" is not.

Rule fields are the hard edges

Banned words are passed as an explicit exclusion — never use these words — and this is the field that most reliably changes output. Keywords work the other way: the terms and facts that should keep appearing, which also feed hashtag scoring. Your emoji policy, hashtag ceiling, CTA style and claims policy sit here too. A claims policy matters more in India than teams expect: if you sell supplements, cosmetics or anything financial, writing "no health outcome claims, no returns figures, no superlatives" into the record is cheaper than editing it out of forty drafts.

Visual and channel fields

Colour palette, fonts, logo and image style condition generated images so a month of posts looks like one brand. The platforms list decides which of the eight networks get slots at all — each one then renders at its own native dimensions and character limit.

Below is a complete record for an invented example — Ember & Chaff, a small Indian roastery selling 250g bags direct. It is not a customer. It is the level of specificity that makes a difference.

With an empty record, a request for a launch caption returns something like this:

Start your morning right with our artisanal, premium coffee blends. Elevate your coffee journey with beans roasted to perfection. Link in bio!

With the record above, the same request returns something like this:

Chikmagalur medium roast, roasted 28 August, shipped 29 August. Filter grind on request. 250g at ₹449.

Nothing changed in the prompt. The second caption is what the constraints allow. Three banned words and one sentence-length rule did most of that work.

A half-filled Brand Brain is worse than an obviously empty one, because the output looks conditioned and isn't. The pattern is always the same: weak values describe a feeling, strong values state a rule someone could check.

Two practical rules. First, if a value would fit any competitor in your category, it is not doing anything — rewrite it until it only fits you. Second, put your last three rounds of editing feedback into the record. Every time you delete a word from a draft, that word belongs in banned words. The record should get sharper each month, and the amount of editing you do should fall.

Each brand carries its own Brand Brain and its own connected accounts. For an agency running eight clients, that is the whole point: the record for a fintech client cannot bleed into a jewellery client's captions, because the two generations read different documents. Brand counts are what the plan tiers are built around — one brand on Free and Starter, three on Pro, ten on Business.

It conditions generation. That is the whole job, and it is worth being clear about the edges.

• It does not verify your claims. If you write a number into the record, the model will repeat it. The claims policy constrains style, not truth.
• It does not learn from performance. Analytics readback is on the roadmap and labelled SOON, along with trend detection, SEO, rendered video and competitor tracking. Nothing today feeds engagement data back into the record — you edit it yourself.
• It does not moderate comments or manage an inbox. Neither is built.
• It does not replace the approval gate. Approval is on by default in autonomous mode, and the daily post cap is enforced by the worker, which counts today's AI posts and stops. The monthly spend cap is stored and displayed, but not enforced by the worker — watch it yourself.
• It does not depend on one AI vendor. The record is provider-agnostic and is passed to whichever of OpenAI, Claude, Gemini, Groq, Mistral, DeepSeek or HuggingFace your workspace is set to.

Twenty minutes, in this sequence. Each step changes more output than the one after it, so if you stop early you have still got most of the value.

1. 01Banned words. List every word you have deleted from a draft in the last month. Ten entries is a good start.
2. 02Writing style. Give a reading level and a maximum sentence length. Numbers, not adjectives.
3. 03Target audience. One sentence with an age range, a city or tier, and what they already own or buy.
4. 04Products. Real names, real prices, in rupees.
5. 05CTA style. Write the exact shape of your ideal CTA and let the model copy it.
6. 06Tone and personality. Three words each, and at least one thing you are not.
7. 07Keywords. The terms and facts that must keep appearing.
8. 08Image style and palette, if you generate images.
9. 09Everything else — mission, vision, goals, competitors, fonts — whenever you have time.

Then generate five captions, and treat every edit you make as a missing field. The Brand Brain page is where you add it.