How do I give an AI my company's voice?
Write a one-page voice guide, do, never, examples, store it in the knowledge base, and include it in every drafting prompt. Then review the first fifty drafts and feed rejections back. Voice is learned from corrections, not descriptions.
The short answer
- Write a one-page voice guide with three parts: things to do, things to never do, and real example sentences, then store it where every drafting prompt can pull from it rather than rewriting it each time.
- The example sentences matter more than the adjectives. Ten real sentences you would actually send give a model specific word choices and sentence structure to pattern-match; 'friendly and professional' describes a goal, not a voice, and gives the model nothing concrete to imitate.
- Voice is learned from corrections, not from a better description. Review the first fifty drafts closely, and when a correction pattern repeats, add a specific rule to the guide rather than assuming the model will improve on its own with more examples of the same kind.
- A voice guide is not a one-time document. As corrections surface new patterns, the guide should grow, and the drafting prompt should always reference the current version rather than a snapshot from when the agent was first built.
Why adjectives fail and examples work
Almost every company, asked to describe its voice, reaches for the same handful of words: friendly, professional, clear, confident. Those words are not wrong, but they are not useful either, because they apply equally well to writing that sounds nothing alike. A model asked to write in a "friendly, professional" voice has to guess at the specifics, sentence length, how formal the greeting is, whether contractions are used, and different guesses land in different places.
Real sentences remove the guessing. Ten actual emails or messages you would send, pasted into the guide as examples, show the model exactly what "friendly" means in your specific case: maybe it means short sentences and no exclamation points, maybe it means a warm opener but a direct ask. The model can match a pattern it can see far more reliably than it can match a description it has to interpret.
What goes into a voice guide
| Input | Where it lives | How it is applied |
|---|---|---|
| Do rules | The one-page voice guide | Referenced in every drafting prompt as explicit instructions |
| Never rules | The one-page voice guide | Treated as hard constraints, not stylistic preferences |
| Ten real example sentences | The one-page voice guide, quoted verbatim | Given to the model as the pattern to match, not only described |
| Corrections from the first fifty drafts | Fed back into the guide as new rules or examples | Closes the loop; repeated correction patterns become explicit rules |
Closing the loop after the guide ships
A voice guide written once and never revisited will drift out of date as the model produces edge cases nobody anticipated. The fix is a review habit, not a better initial guide: read the first fifty drafts closely, note what gets corrected, and look for patterns rather than treating each correction as a one-off. If three different drafts all get the greeting corrected the same way, that is worth a rule in the guide, not three separate manual fixes repeated indefinitely.
Disclosure: SalesCrew is our product, and its drafting features are built to pull from a knowledge base entry holding the client's voice guide, included in every relevant drafting prompt so the reference stays current without editing each agent individually. Draft review, through the approval queue, is where corrections happen and where a pattern worth adding to the guide gets noticed.
'Friendly and professional' is not a voice; ten real sentences are
Questions
- Why is 'friendly and professional' not a voice guide?
- Because it describes almost every company's stated goal for its writing, and it gives a model nothing concrete to imitate. Ten real sentences you would actually send show specific word choices, sentence length and tone, which is what a model can actually pattern-match against.
- How many drafts should you review before trusting a voice guide?
- A reasonable starting point is the first fifty drafts, reviewed closely enough to catch patterns, not only individual mistakes. If the same kind of correction keeps showing up, that is a signal the voice guide is missing a rule, not that the model needs to try harder.
- Should the voice guide be the same for every kind of message?
- Mostly, but note the exceptions explicitly. A cold outreach opener and a reply to an existing customer's complaint call for different registers even within the same overall voice, and the guide should say so rather than leaving the model to infer it.