Persona and tone

You now know the six parts of a system prompt. This lesson zooms in on the one that people feel most strongly about and describe most badly: the voice. Every business owner who has tried AI has had the moment where the output was technically fine and they still would not post it, because it did not sound like them. The fix is not "try harder." It is knowing what a persona actually controls, and giving the model the four or five specific things it needs.

Persona is a job, tone is how it talks

Two words get mixed up here, so let's separate them. Persona is who the AI is playing: the owner of the shop, a patient teacher, a seasoned irrigation tech. It sets what the model knows, what it cares about, and what it assumes the reader knows. Tone is the texture of the words: warm or crisp, playful or plain, short or flowing. You can have the same persona in several tones (a friendly owner, a tired owner, a formal owner) and the same tone across personas.

Most people only set tone ("be friendly") and skip persona. Then they wonder why the friendly output still sounds like a chatbot. It is because a friendly nobody is still a nobody. Set the persona first, then tune the tone.

Tap between the three personas below. The question is identical. Watch how the answer changes not just in wording but in what it decides to mention, what it skips, and what it assumes you already know.

The five dials

When you write the persona and tone section of a system prompt, these five settings cover nearly everything. Set the ones you care about and leave the rest.

• Voice. Who is speaking, in one line. "You speak as the owner, first person, like a neighbor over the fence." Or: "You write as the company, never as a person, and never use the word I."
• Reading level. The single most effective tone instruction, and almost nobody uses it. "Sixth-grade reading level" strips out the fancy vocabulary and long sentences that make AI text feel stiff. Most customer-facing writing for a Southern Utah business should land between sixth and eighth grade. Internal notes for yourself can go higher.
• Warmth. Plain versus personable. "Warm but not gushing" is fine as a starting point, but attach a rule to it that the model can actually follow: "No exclamation points. Thank people once, not three times."
• Personality. How much of a character is allowed. Zero is fine. A little humor is fine. What does not work is "be funny," because the model's idea of funny is a dad joke about lawns. If you want humor, give it a sample of the kind you like.
• Formality. Contractions or not, first names or last names, "hey" or "hello." Small choices, but they are exactly the things a reader notices.

Matching a brand you already have

If you have been in business a while, you already have a voice. It is in your website, your Google reviews replies, your best Facebook post, the text messages you send to customers. Your job is not to invent a persona. It is to describe the one you already have, precisely enough that a model can reproduce it.

Start by collecting three pieces of writing you are genuinely proud of. Not the ones that got the most likes, the ones that sound most like you. Then read them and write down what you notice. Short sentences? Do you name the customer? Do you mention the weather, the town, the crew? Do you ever apologize, and how? Do you sign off, and how? Those observations become your tone rules.

You can also ask the AI to do this analysis for you. Paste the three samples in a normal chat and say: "Describe the voice in these three pieces in ten short rules a writer could follow." It is surprisingly good at noticing patterns you have stopped seeing in your own writing. Edit its list, keep what rings true, and drop that into the system prompt.

Sample-based tone: the one that actually works

Every dial above helps. But the single most powerful move is one you already learned: paste a sample. A system prompt that ends with "Match this voice:" and two real sentences beats one that spends a paragraph on adjectives. Adjectives are opinions. A sample is data, and the model is built to continue data.

Pick samples carefully. They should be short (two or three sentences), typical (not your one brilliant post), and recent (your voice five years ago may not be your voice now). If your business writes in more than one register, say, playful on Facebook and straight in a quote email, include one sample of each and label them: "Facebook voice:" and "Email voice:". The model will pick the right one based on what you ask for.

A caution about copying someone else. It is tempting to paste a competitor's post you admire. Do not. You will get a competent imitation of them, and your customers will hear it. The point of a persona is that it is yours.

Words to ban. One more tone tool that punches above its weight: a short list of words and phrases the model may not use. Every model has verbal tics, and they are the first thing readers notice. If your output keeps saying "elevate," "seamless," "we've got you covered," or opening every post with "Attention homeowners," a banned-words line fixes it in one stroke.

Build the list from real output, not from guesses. Use the assistant for a week, and every time a phrase makes you wince, add it. Ten to fifteen entries is plenty. Write it as a plain line in the persona section: "Never use these words or phrases: [list]." It is the fastest tone improvement most people ever make, and it is specific enough that the model actually follows it.

When the persona should not be you

Not every system prompt should sound like the owner. Sometimes the right persona is the expert you wish you had on staff. A landscaper might keep a second prompt where the AI plays a plant pathologist for diagnosing yellowing leaves. An insurance agent might want one that plays a careful compliance reviewer. A restaurant owner might want a blunt food critic for menu feedback.

These expert personas follow the same rules, with one addition: name the expertise specifically and tell the model how sure to sound. "You are an experienced irrigation technician. When you are not certain, say so and suggest what to check." Without that last sentence, an expert persona tends to sound confident about everything, including things it is guessing at. Lesson 4 comes back to this.

The other place to break from your own voice is reading level for yourself. When the AI is writing for your customers, aim low. When it is explaining something to you, let it use the real words. A quick rule in the prompt handles it: "Customer-facing text: sixth grade. Explanations to me: normal adult."

Try this yourself

Do the sample-collection exercise now, before you move on. Open a normal chat (no system prompt yet) and paste this, replacing the brackets with three real pieces of your own writing.

Here are three things I wrote for my business. Read them and describe my writing voice as ten short rules that another writer could follow exactly. Cover sentence length, formality, how I address the customer, what I tend to mention, and what I never do. Then pick the two sentences from these samples that best represent the voice.

Sample 1:
[paste]

Sample 2:
[paste]

Sample 3:
[paste]

Take the ten rules, cross out any that feel wrong, and paste the survivors plus the two chosen sentences into the PERSONA and EXAMPLE sections of the template from lesson 2. That is your voice, written down for the first time. Most owners find that part strangely satisfying.