How to actually talk to it.

Most people get mediocre results because they ask like it's Google. Four habits fix that, and you can learn them in an hour.

Stop searching, start briefing

The instinct everyone has on day one is to type a short question, the way you'd type into Google. "write a follow up email". You get back something generic, you conclude AI is overhyped, and you go back to writing it yourself.

The fix is a change of posture. You are not searching. You are briefing a capable new hire who is fast, widely read, and has never seen your business. Everything below follows from that.

1. Tell it who it is and who it's for

One sentence of role and audience changes the output more than any other single thing.

You're the office manager at a three-dentist practice in Austin. Write a text to a patient who missed a cleaning, offering two times this week. Warm, short, no guilt.

Compare that to "write a text about a missed appointment." Same model, completely different result — because the second one has no idea whether you're a hospital or a barber.

2. Give it the raw material

The most common mistake is describing your information instead of providing it. Don't summarise the email — paste the email. Don't describe the spreadsheet — attach it. These tools handle enormous amounts of text now; some read a million words at once. You are not going to overwhelm it.

If your answer depends on something the model can't see, it will invent a plausible version. That's not the model lying to you — it's you asking a question that can't be answered from what's on the table.

3. Show it one good example

This is the highest-leverage habit in the whole list, and almost nobody does it. If you want output that sounds like you, paste something you already wrote that you liked.

Here are two proposals I've sent that won the work. Match this voice — short paragraphs, no jargon, price stated plainly without apology. [paste proposal one] [paste proposal two] Now draft one for a client who needs [job], budget around [amount].

One example beats a paragraph of adjectives describing the tone you want. Three examples beats one.

4. Tell it what to do when it doesn't know

Left to itself, a model will fill a gap with something that reads well. Give it a different instruction and it will take it.

If anything is missing that you'd need to answer properly, ask me before writing. Don't guess at numbers, dates, or names.

Add that line to anything that matters. It converts confident nonsense into a short list of questions, which is what you wanted.

The part people skip: go again

The first output is a draft, not a delivery. The single biggest difference between people who get value out of this and people who don't is that the second group stops at attempt one.

Say what's wrong in plain language — "too formal, cut it in half, lose the last paragraph" — and it will fix it. Three rounds takes ninety seconds and gets you something you'd actually send.

Save what works.

When a prompt finally produces the right thing, paste it into a note. In a month you'll have a dozen you reuse constantly, and that file is worth more than any course. That's what our snippet library is — ours, after two years.

Do this now, with something real

Pick a task from this week you actually have to do — an email, a summary, a plan.
Write four lines: who it's for, the raw material, one example you liked, and "ask me if something's missing."
Push back three times before you accept the answer.
Save the prompt if it worked.

Where to go next

Once this feels natural, the next question is when to stop chatting and build something that runs on its own. That's agent vs. a single call — six minutes, and it saves people from building the wrong thing.

Want someone to just do this with you?

The AI Once-Over maps where AI actually saves time or makes money in your business, then builds the first one with you.

See the AI Once-Over →