The brand tool: beliefs can't be written by AI, so we surface them

The problem behind it

YuktiBOS had a field called "Brand principles". When I looked at what was in it, I found colours, fonts, design tokens and one paragraph about voice. Nothing said what the client believes, who it is for and not for, how it behaves, or what it never says.

I wanted a brand tool that sits in a fixed order with the others. The order I proposed on 31 October: Market finder, then Offer and Pricing, then Brand, then Voice and tone, then the Design system, then content and ads. Each feeds the next.

What we learned from the research first

Before anyone wrote a spec, a research note compared the frameworks: Keller, Aaker, Unilever's Brand Key, Kapferer, Neumeier, Sinek, the twelve archetypes, and Moore's positioning sentence. Its short answer:

• Every serious framework asks the same few questions: who it is for, what they would use instead, what only we do, why anyone should believe us, and what character we show.
• Beliefs only matter when they cost something. The note quotes Lencioni: values properly practised inflict pain, and Enron's values were "Communication. Respect. Integrity. Excellence."
• Airbnb's purpose came from listening: 485 interviews in three and a half weeks.
• Nothing on the market does what we wanted. Brand portals host rules someone else wrote. AI writing tools capture voice from samples. Nobody surfaces beliefs from the founder's own words with the source shown.
• Small businesses do not need a brand book. They need it used. In one study almost every firm had guidelines and only about a quarter enforced them. In YuktiBOS agents read the brand on every task, which solves enforcement by design.

The key decision: surfaced, never written

My line on 31 October: AI can write, but beliefs AI cannot write. Ninety-nine percent of clients will not have written beliefs. If they did, they would already have brand principles.

So the rule became:

• Yukti finds candidate statements in the person's own talk and writing.
• The person swipes Agree, Close or Disagree.
• Each belief carries "You said: ..." with the quote. A belief without a quote from the person is marked Suggested and needs an explicit Keep.
• Yukti never invents a belief.

In the build, code enforces it. A quote is checked against the source text, and only the server can mark it verified. Three of the twelve swipe cards are common views from the field. Agreeing with one keeps it Suggested until the person writes a one-line reason. Imported beliefs need one Keep each, never pre-kept.

What a brand system contains

Six parts: a position line (one master line plus a line for each audience, "the room dictates"), four to six beliefs in a fixed form (we believe, so we, even when), the audience, personality, proof, and red lines. No colours and no fonts. A lint refuses them in every save.

Checks run in code: the position line is 25 words or fewer; it passes a competitor-swap test (put the main alternative's name in its place; if it still fits, it fails); it carries one benefit. Legal and personal red lines need an explicit Yes, stamped with the person's name and the time.

Other decisions

• Never ask twice. The tool reads the interview, the website, the chosen market, the picked offer, the voice guide and Client knowledge first. It asks only for gaps.
• Cost by design. Opus for belief mining and positioning, Sonnet for drafting, Haiku for extraction. A hard cap of Rs 150 per build: the worst case of each call is estimated before it starts and refused if it would pass the cap.
• Moving the old field safely. Colour, font and token lines moved to the top of the Design system. The voice paragraph became a voice source. The old text is kept in the activity log. Nothing deleted.
• Archetypes are a hint, never a box. The evidence for them is weak.

What went wrong

On 31 October the live review found a dead end: the "Build my brand system" button was wiped on first load. React runs a child's effect before its parent's, and a reset effect in the shell cleared the label the child had just set. We fixed it by tagging overrides with the part that set them. The voice shell had the same latent bug.

A client-rendered website returned only 23 words to our reader. The reader now also pulls meta tags, structured data, the sitemap and up to five key pages, and warns under 300 words.

What's next

The brand finder is built (ADR-344). The design system step that follows it is, from my own notes, a known weak spot: the Mukika design system exists on my Mac but the app holds only a scrape of the code. The Pricing tool needed review. I have not verified a finished end-to-end run on a second client.

Numbers at a glance