You write a sharp brief, hit enter, and the draft comes back sounding like every other company on the internet. Friendly, vague, and weirdly proud of itself. Run it again tomorrow and the tone shifts. Run it in a different tool and it shifts again. Most people blame the model. The real cause is that nobody told the model what “you” sounds like, so it falls back on the average of everything it has read.
What this post covers: A consistent AI brand voice is something you build as a system, not something you hope for. This post is for founders, creative directors, and agency operators who already use AI daily. You will learn how to write a short voice guide the AI can read, which words to ban, how to test the result, and how to carry the same voice across every tool.
Table of Contents
Why AI Output Drifts Away From Your Voice
AI output drifts because a language model without guidance writes toward the statistical middle. It has no memory of your last post, no sense of your clients, and no opinion about how you talk. It gives you the safest version of a sentence.
That safe version has a recognizable shape. It opens with a warm greeting, lists three benefits, and closes with a line about the future. Wikipedia’s editors catalogued these habits in their 2025 “Signs of AI writing” guide, including inflated importance, promotional adjectives, and tidy groups of three. Once you know the list, you see it everywhere.
Drift also happens inside one project. A prompt written on Monday and a prompt written on Thursday are never quite the same, so the tone moves with your mood. Add a second tool and a teammate who prompts differently, and you have three voices pretending to be one brand.
I see this constantly when I set up workflows for agencies. The team has a brand book with colors, logo rules, and a paragraph about “being approachable.” None of it is written in a form an AI can act on. Adjectives like “approachable” give the model nothing to copy. Examples and rules give it everything.
The fix is to treat voice like any other asset in your stack. If you want the bigger picture on how these pieces connect, the AI Orchestra workflow shows where a voice file sits next to briefs, tools, and review steps.
Write a Voice Guide the AI Can Read
A voice guide the AI can read is a one-page file made of real examples and plain rules, not personality adjectives. Keep it under 600 words so it fits comfortably at the top of any prompt.
Start from your own writing, not from a blank page. Several vendor guides, including StoryChief’s walkthrough for building an AI brand voice, follow the same pattern: feed the tool samples of your real content and let it describe your tone and rhythm. That is the right order. You are the source. The AI is the note-taker.
Here is the process I use.
Step 1. Collect samples. Pick three to five pieces you are proud of. Emails, posts, a proposal, a transcript of you explaining something out loud. Variety matters because it shows what stays constant when the format changes.
Step 2. Ask the AI to describe the patterns. Paste the samples and ask for sentence length, how you open, how you handle opinions, what you never say, and what you repeat. Read the answer critically. It will flatter you. Cut anything that sounds like a compliment and keep anything you can verify in the samples.
Step 3. Write the rules. Turn the patterns into short instructions. “Open with the point, not a greeting.” “Use first person when sharing a result.” “Name the tool or the number, never ‘various tools’.” Rules beat adjectives because a rule can be checked.
Step 4. Test and edit. Give the guide to the AI with three old briefs you already wrote by hand. Compare. Wherever the output feels off, the guide is missing a rule. Add it and run again.
Add two more blocks before you finish. First, a short list of channel shifts: how you sound on LinkedIn versus in a client email versus in a proposal. Second, three to five “gold” sentences you wrote that sound exactly like you. Models imitate examples far better than they follow descriptions.
Words and Patterns to Ban
A ban list works because it is the fastest way to remove the habits that make AI writing obvious. It takes ten minutes to write and pays back on every output.
Start with the words. Make your own list, but the usual offenders are inflated verbs and adjectives that sound impressive and say nothing: “world-class,” “revolutionary,” “supercharge,” “thrilled,” “dive into.” If a word would feel odd coming out of your mouth in a meeting, it goes on the list.
Then ban structures, which matter more than single words:
- Em dashes used as punctuation tics. Swap them for a period or a comma.
- Rule-of-three lists where two items would do. AI loves triplets because they sound complete.
- Negative parallelism, the “it is not just X, it is Y” move.
- Vague attribution, such as “experts say” with no named source.
- Chatbot openers and generic closers, like “In today’s fast-paced world” or “the future is bright.”
Put the list in the voice guide as a plain instruction: “Never use these words or patterns.” Then add a final pass. After the draft is written, ask the AI to scan its own output against the ban list and rewrite any hit. I run my own posts through a humanizer step for exactly this reason, and it catches things I would miss on a tired read.
A ban list is not a substitute for taste. It cleans the surface. The deeper work is deciding what you actually think, which is the editing job no model can do for you. The blog archive has more on that, because it is where most AI-assisted writing goes wrong.
Before and After: One Paragraph, Two Voices
The clearest way to see the effect is to give the same brief to the same model with and without a voice guide. The brief: announce a new client onboarding template. The numbers in the second version are illustrative.
Without a guide:
We are thrilled to announce a world-class onboarding template that helps teams supercharge their workflows and deliver exceptional client experiences in today’s fast-paced world.
With a guide (short sentences, first person, named specifics, no hype):
I rebuilt our client onboarding template last month. Revision rounds dropped from four to two on the next three projects. It is a single page. Here it is.
Look at what changed. The second version names a number, uses first person, keeps sentences short, and ends on an action. None of that came from a clever prompt. It came from rules the AI could follow.
Notice also what the guide did not do. It did not make the AI smarter. It narrowed the options so the model stopped picking the safest phrase. That is the useful way to think about voice: constraints, not creativity.
Keep a Consistent AI Brand Voice Across Tools
You keep a consistent AI brand voice across tools by storing the guide in one place and loading it into each tool the same way. The mistake is rewriting the voice separately inside every app.
Make one master file. I keep mine as a plain markdown file called voice.md. Then I load it where each tool allows it:
- Claude: paste it into a Project’s instructions so every chat in that project starts with it.
- ChatGPT: add the key rules to custom instructions, or attach the file in a custom GPT.
- Gemini: use a saved Gem with the guide as its instruction block.
- Notion AI and other embedded assistants: paste a condensed version, since these often have short instruction limits.
The condensed version matters. Most tools have a character cap on instructions, so keep a 150-word summary of the guide: five rules, the ban list, and two gold sentences. The full file lives in your master copy, and the summary travels.
Then schedule maintenance. Once a month, take three outputs from the past weeks and check them against the guide. Where the voice slipped, fix the file, not the individual draft. A voice guide that never changes is probably being ignored.
If a teammate prompts the tool too, they get the same file. That is the part agencies tend to miss. A voice system only holds when it is shared, which is why the about page of my practice describes the work as setting up AI workflows for teams and not just writing prompts. Consistency comes from the system around the model.
Finally, keep a human read in the loop. Dotdigital’s guide to building a brand voice guide that AI can actually use makes the same point: write it for the machine, with real examples and constraints instead of adjectives. Even a well-built guide produces the occasional miss, and you are the one who spots it.
Key Takeaways
- AI drifts into a generic tone because it writes toward the average when you give it nothing specific to copy.
- A usable voice guide is under 600 words and made of rules and real examples, not personality adjectives.
- Build the guide from your own writing: collect three to five samples, have the AI describe the patterns, then keep only what you can verify.
- Ban structures as well as words: em dash tics, forced triplets, negative parallelism, vague attribution, and chatbot openers.
- Test the guide against three briefs you already wrote by hand, and add a rule wherever the output feels off.
- Store one master file and load it into every tool, with a 150-word summary for tools that limit instruction length.
- Review three outputs a month and fix the file, not the single draft.
Frequently Asked Questions
How long should an AI brand voice guide be?
Aim for 400 to 600 words. Long enough to include rules, a ban list, channel notes, and a few example sentences, short enough to sit at the top of a prompt without crowding out the actual task. If it grows past a page, split it into a core file and a channel-specific file.
Can AI really learn my writing style?
It can imitate patterns it is shown, such as sentence length, openings, and vocabulary. It does not absorb your opinions or judgment. Give it real samples and explicit rules, then edit the result. Expect strong surface imitation and a need for your review on substance.
Why does my AI voice change between tools?
Each tool starts from its own defaults and its own instruction limits. If you wrote the voice separately in each one, the versions diverge. Keep one master file and load a condensed copy into every tool so the source stays the same.
What words should I ban in my AI voice guide?
Ban the inflated words you would never say aloud, such as “world-class,” “revolutionary,” and “supercharge.” More important, ban patterns like forced lists of three, em dash tics, and vague attribution. Build the list from your own annoyance with past drafts.
How often should I update my voice guide?
Review it monthly. Pull three recent outputs, compare them to the guide, and add a rule wherever the tone slipped. Update it sooner when your positioning, audience, or channels change.
Harshal Saraf is a Creative Director and AI Workflow Consultant based in Indore, India. Under his practice ByHarshal, he sets up AI workflows for founders, agencies, and brands across India. Where Creative Direction Meets AI Orchestration. He has led creative direction for brands and small and medium scale B2B businesses, and currently works as Creative Director and AI Strategist at Square Root SEO. He writes Oh, So AI, a Tuesday and Friday newsletter on AI tools, workflows, and productivity for founders and creatives.
