Writing an AI persona that actually sounds like you

Most AI replies fail for the same reason, and it is not the model. It is that the persona describes a role instead of a voice.

Here is the prompt almost everyone writes first:

You are a helpful, professional assistant for a freelance designer. Respond politely and courteously to client enquiries. Always be friendly and maintain a professional tone.

Every word of that is true and none of it is useful. "Professional and courteous" is the default behaviour of every language model on earth — you have described the thing it already does. So you get the thing it already does: complete sentences, slightly formal register, an offer to help further. Perfectly pleasant, and instantly recognisable as not you.

Describe mechanics, not attributes

The fix is to stop writing adjectives and start writing observable rules. "Casual" is an adjective. "I don't use full stops in one-line replies" is a mechanic — it produces a visible difference in the output.

Open your own WhatsApp, find five replies you sent to clients, and look for what is physically true about them:

• Length. How many lines is your typical reply? Most people write far shorter than the model does by default.
• Greetings. Do you open with "Hi Sarah," or go straight into the answer? Most people in an ongoing thread go straight in.
• Punctuation and capitals. Sentence case or lowercase? Full stops on short messages? Em dashes, or commas?
• Contractions and filler. "I will" or "I'll"? Do you start sentences with "so" or "yeah"?
• Sign-offs. Do you have one? Most people do not, in chat.
• Your actual stock phrases. Everyone has four or five. "Let me check and come back to you." "Happy to take a look." Write them down verbatim.

That list is your persona. It is much less elegant than a paragraph about professionalism, and it works enormously better.

A rewrite

Same freelance designer, written as mechanics:

You reply as me, a freelance brand designer. Write like WhatsApp, not email: usually one to three short lines, no greeting when the thread is already going, no sign-off. Sentence case. Contractions always. It's fine to start a sentence with "so" or "yeah". Skip the full stop on one-line replies.

Answer the question first, then add at most one useful detail. Never list options with bullet points. If someone asks about price or timeline, quote the numbers from the knowledge base exactly — never estimate or round. If it isn't in the knowledge base, say you'll check and come back rather than guessing.

Phrases I actually use: "happy to take a look", "let me check and come back to you", "that works on my end". Don't use "certainly", "I'd be delighted", or "reach out".

That last paragraph — the explicit banned words — does a surprising amount of work. Models drift toward a small set of customer-service phrases, and naming them is more effective than asking for a different tone in the abstract.

Separate voice from facts

A persona should describe how you write. It should not contain your prices.

The reason is practical: facts change and voices do not. If your day rate lives in the persona paragraph, you will edit your voice every time you raise it, and the model will treat the number as a soft suggestion it can paraphrase. Facts belong in a knowledge base — quoted exactly, not rewritten.

In auto-reply.to these are two separate inputs for exactly this reason: the persona governs style, the knowledge base holds the things that must come back verbatim. It also means the knowledge base can be encrypted at rest while the persona is just prose.

Get the deferral right

The most important behaviour to specify is what happens when the assistant does not know something.

Left to its own devices, a model will produce a plausible answer, because producing answers is what it does. That is fine for trivia and catastrophic for a quoted delivery date. So write the deferral explicitly, in your own words:

If you don't have the answer in the knowledge base, don't estimate. Say something like "let me check and come back to you on that" and stop. Never guess at prices, dates, or availability.

Then test it deliberately. Ask your assistant something you know you did not cover and confirm it defers instead of improvising. If it invents an answer, the persona is not done. This is the single check most worth doing before you whitelist a second chat, and it is worth repeating whenever you edit the persona.

Tune from the transcript, not from memory

Once it is live, resist the urge to rewrite the persona from a general feeling that replies are "a bit off". Find the specific reply that was wrong, work out which rule was missing, and add that one rule.

If you can see why a reply was written — the reasoning behind it, not just the output — this gets much faster, because you are fixing a cause rather than guessing at one. That is what the audit log is for.

A rhythm that works: read every reply for the first week, add one rule at a time, and stop when replies stop surprising you. Four or five iterations is normal. You are not writing a document, you are converging on a voice.

Where to go next

Set the persona up before you connect anything — it takes twenty minutes and it determines everything downstream. If you are still deciding whether to automate at all, the safety trade-offs are worth reading first. If you want to see how the persona differs by line of work, the use cases have concrete examples, and the trial lets you test yours on one conversation.