How to Keep AI-Drafted Posts Sounding Like Your Brand
Turning Voice From an Adjective List Into Instructions a Model Can Follow
Short answer
AI drafts sound generic because voice was described in adjectives rather than rules. Replace words like friendly and bold with constraints you can count or search for, give the model five to ten of your own published posts as examples, and check every draft against one test per rule.
1. What makes a brand voice something an AI can actually follow?
A followable voice is a short list of testable constraints, not adjectives. Each rule names something a reader can count or search for in a draft, which is what makes a violation visible to a reviewer and to a model.
- Brand voice
- The consistent set of language choices that makes a company recognisable in writing: sentence rhythm, vocabulary, person, humour and the way claims are made.
Adjectives fail because two people read them differently. Friendly means emoji to one writer and first names to another, and a model trained on most of the public web reads it as the average of every friendly brand it has seen. That average is the generic voice you are trying to escape.
| Attribute | What the rule says | How to test a draft against it |
|---|---|---|
| Sentence length | Most sentences under 20 words, at most one subordinate clause | Find the longest sentence and rewrite anything over 25 words |
| Person | We for the company, you for the reader, never one or businesses | Search for businesses, companies and one used as sentence subjects |
| Claim style | Every number carries its basis, no superlative without evidence | Circle each adjective and delete any that no number supports |
| Vocabulary | Plain words instead of leverage, solution, ecosystem, unlock | Read it aloud and flag any word a customer would not say |
| Openers | First line states the point, never a rhetorical question | Check whether line one would work as the entire post |
| Humour | Dry, aimed at the work or at us, never at the reader | Ask whether the person in the story would laugh at it |
2. Why do AI-drafted posts sound generic by default?
A model with no voice instructions writes toward the average of its training data, which is the middle of your category. Underspecified prompts produce competent copy any competitor could publish unchanged.
The generic register has recognisable tells: a rhetorical question as the opening line, a three-item list where two items would do, the word solution standing in for the product, and a closing sentence inviting the reader to share their thoughts. None of it is wrong. All of it is interchangeable.
Prompt length is not the fix. A thousand words of aspirational brand language produce the same output as fifty, because none of it constrains a specific choice. Ten lines that can each be checked will change more than a page of description.
The second cause is source material. A model given no examples writes from its memory of the category. A model given your last ten published posts has something concrete to imitate, and imitation is what these systems are genuinely good at.
3. How do I write voice instructions a model will follow?
Convert each voice trait into a rule that can be checked, add explicit do-not rules with replacements, and attach real examples. Instructions naming a countable property change output. Instructions naming a feeling do not.
This takes an afternoon once, and the output is a page you paste into every drafting prompt.
- Collect ten posts you would publish again. Pull the most on-voice posts you published in the last quarter. Ten is enough. They become both the examples the model imitates and the evidence for every rule you are about to write down.If the ten have nothing in common, the voice does not exist yet, and no prompt will invent one.
- Extract the pattern rather than inventing it. Read the ten side by side and note what they share: how they open, typical sentence length, whether they ask questions, which words repeat. Write down the voice you already have, not the one you would like after a rebrand.
- Turn each pattern into a countable rule. Rewrite every observation as something a reviewer can verify in seconds. Opens with the point. Sentences under 20 words. No adjective without a number behind it. Aim for six to ten rules in total.
- Add the do-not list with replacements. List the words and moves you never use, and give each one a substitute. Deny rules are easier for a model to obey than positive ones, because they name a specific token instead of a target it has to infer.
- Include three counter-examples. Add two or three drafts that are off-voice, each with one line explaining why it fails. A model shown the boundary makes fewer boundary mistakes than one shown only the centre of the target.
4. How do I test whether a draft is on-voice?
Run one check per rule, in order, and stop at the first failure. A rule-by-rule pass takes under a minute per post and produces a specific fix instead of a reviewer saying the draft feels off.
Five checks, in this order, because the early ones fail fastest:
- Read the first line alone. If it does not carry the point, nothing after it matters.
- Search for the words on your deny list. This is the quickest check and it catches the most drafts.
- Count the longest sentence. One long sentence is a choice, three in a row is the model default.
- Check every claim for its basis. A number with no source and a superlative with no number are the two claims that cost trust fastest.
- Read it aloud, one breath per sentence. Anything you stumble over is not how the brand talks.
The last check is the swap test: replace your company name with a competitor's and read it again. If the post still works unchanged, it says nothing only you could say, and that is a positioning problem the voice rules cannot fix.
5. Should the voice change per channel, and how does it stay current?
Keep one voice and vary the register per channel. Format, length and formality change; the deny list, the claim style and the person do not. Re-extract the rules from recent posts once a quarter.
A LinkedIn post and an Instagram caption from the same company should read like the same writer in two different rooms. The rules that travel unchanged are the deny list, the claim style and the person. The rules that flex are length, formatting, emoji and how much context you restate.
Drift is the real failure mode. Rules written in January describe a company that has since changed its pricing, its main objection and part of its vocabulary. Run the extraction step again each quarter on the last ten posts you were happy with, and update the rules where the pattern moved.
The rules only work if they reach the person drafting. Oppira keeps voice rules and example posts in the Brand Kit that every generated draft is written against, and it names the fields still left empty, which is usually where the generic sentences come from.
Key Takeaways
Adjectives are not instructions
Friendly and bold mean something different to every writer and to the model. Only rules you can count or search for change the output.
Extract the voice you have, not the one you want
Ten published posts you would run again contain the real pattern. Write the rules from those rather than from a brand workshop.
Deny lists work harder than positive rules
A list of words you never use, each with a replacement, is easier for a model to obey than a description of a target tone.
Test one rule at a time
A rule-by-rule pass takes under a minute and returns a specific fix, instead of a reviewer reporting that the draft feels off.
Apply the swap test before publishing
Put a competitor name where yours is. If the post still works, it says nothing only your company could say.
Re-extract the rules every quarter
Voice rules describe the company that wrote them. Refresh them from recent posts or they slowly stop matching how you talk.
Frequently Asked Questions
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. Start free and see your market clearly by tomorrow.