How to Make ChatGPT or Claude Sound Like Your Brand

How to make ChatGPT and Claude sound like your brand — voice consistency guide

The "AI sounds generic" complaint is real, fixable, and almost never about the AI. It's about what you fed it. Six concrete steps that move every AI tool from "sounds like a chatbot template" to "sounds like our senior copywriter." Part of our AI workflows pillar. For the lasting fix: Branva Brain.

The customer pain point

Founders try AI for copy. The first 2-3 prompts look promising. By the 20th the replies all blur into the same generic "We understand your frustration. Let me help you with that." tone. They conclude AI can't capture brand voice and go back to writing by hand at 11pm.

The diagnosis is wrong. AI didn't fail to capture brand voice — you didn't actually give it the voice. You gave it a vague prompt and three adjectives. That's not voice. Adjectives describe a voice; they aren't one.

This post is the fix — the right way to feed voice to AI, with each step's failure mode honestly called out so you know when to invest in the next layer.

Table of Contents

Why "be friendly and witty" doesn't work

The single biggest reason AI sounds generic: people tell it to be friendly. "Be warm but witty. Casual but professional. Direct but empathetic."

Every adjective in that sentence is also every other DTC brand's adjective. The model has no signal to distinguish you. So it defaults to the safest middle of the friendly-witty-casual distribution — which is exactly the chatbot-template tone you wanted to escape.

Adjectives are not voice. Examples are voice. Specific words are voice. Sentence rhythm is voice. The things you don't say are voice.

Step 1 — Give it real examples (not adjectives)

Take 5-10 of your best recent replies, posts, or emails. Real ones. Paste them into the prompt with a one-line annotation each:

Example 1 — Damaged-product reply. "That's a rough first wear, sorry about that. We do replace on quality issues like this within 60 days, no return shipping. Here's the link to start the swap." Notes: Drop the apology fast. Action-first. Specific policy in their own words.

Example 2 — Welcome email opening. "Hey — first order in. We'll send the order confirmation in a sec; this is the version where I tell you how I think about the brand without it being a wall of text." Notes: First-person where natural. Short paragraph rhythm. Light self-awareness.

Three to five real examples beat any number of adjectives. The model learns from the patterns it sees, not from your description of the patterns.

Failure mode: ad-hoc. Every new chat means re-pasting examples. The voice drifts when someone forgets. We'll fix that at step 6.

Step 2 — Write the anti-examples too

What you don't say is half of voice. Write 5 anti-examples — phrases your brand explicitly avoids:

Avoid: "I understand your frustration." Avoid: "We appreciate your patience." Avoid: "Please don't hesitate to reach out." Avoid: Anything starting with "Unfortunately," Avoid: "We hope this helps!"

Then add: "if the draft includes any of these, rewrite."

Anti-examples are what separate a real brand voice from chatbot-friendly default. They give the model a hard signal: not just "sound more like X" but "absolutely don't sound like Y."

Failure mode: still ad-hoc. Same problem as step 1 — it only works in the one chat where you set it up.

Step 3 — Capture how voice shifts by situation

Brand voice isn't monolithic. The voice on a welcome email is different from the voice on a damaged-product response, which is different from the voice on a press inquiry. Document the shifts:

Situation Voice shift
Welcome / first-touch Lighter, slightly self-aware, founder-first-person OK
Standard CS reply Action-first, friendly but direct, no fluff
Damaged product Apologize once, fix immediately, no over-apologizing
VIP customer issue Warmer, name-checked, manager involvement signaled
Press / partnership inquiry Tighter, brand-rep voice, slight remove

Without this matrix, the model averages everything into one mid-tone voice. With it, the right voice lands for the situation.

Failure mode: the matrix is dense and adding it to every prompt is brittle. Same chat-by-chat decay problem.

Step 4 — Include policies, not just tone

Voice is only half the problem. The other half is that AI invents policies that don't exist. Customer asks "do you offer free returns?" — generic AI says yes, 90 days, free shipping. Real policy: 30 days, customer pays return shipping.

You can't have on-brand content without on-policy content. Document the policies and feed them to the model:

Even better — phrase the policies in your voice. "We do free returns in 30 days. We don't pay return shipping; the customer does. If it's a quality issue we eat the shipping — that exception lives in the damaged-product playbook."

Failure mode: policies change. Sales change. New products launch with new shipping rules. Your pasted policy block goes stale. You either re-paste every time something changes, or the AI starts citing the wrong policy.

Step 5 — Document the edge cases

The 15-20% of cases that aren't standard are where AI fails most loudly:

These are where generic AI defaults to the worst possible reply because the patterns it's seen in training don't match your brand's specific handling. The fix: document each edge case with the trigger pattern + the response pattern + the escalation rule.

Failure mode: the edge case list is long. Your senior CS lead knows them all. Getting them out of her head into structured docs is a real project — 2-4 weeks for most brands.

Step 6 — Stop pasting it into every chat

If you've done steps 1-5 well, you now have a real brand voice document. Maybe 5-15 pages. Voice examples, anti-examples, situation matrix, policies, edge cases.

You can paste it into every new Claude or ChatGPT chat. It'll work. But two problems show up at scale:

  1. It's not in your other tools. Your Klaviyo flow drafter doesn't have it. The new freelancer's Perplexity searches don't have it. Each tool needs its own copy.
  2. It goes stale. Policy changes, new products launch, edge cases evolve. Manual re-pasting and re-uploading is a maintenance burden that doesn't get done.

The structural fix is a brand brain: the same voice + policies + edge-case knowledge stored once, kept synced, queryable from every AI tool you use via MCP and from automation via REST. You stop pasting. Every tool reads from the same canon.

This is what Branva Brain is. The full architecture is covered in What is a brand brain?.

The honest end state

Steps 1-5 will materially improve the AI output you're already getting today. That's real. Do them this week.

Steps 1-5 also have the same operational ceiling: they only work in the one chat or one tool where you set them up, and they degrade the moment your policies change. For a growing DTC brand running marketing across 5-7 tools and multiple AI surfaces, that ceiling matters.

The lasting fix is the brain — set up once, kept synced, available everywhere. Until then, the steps above buy you a meaningful improvement and a real understanding of what you'll eventually want to systematize.

Related reading