Most people use AI the way they use a search engine: type a question, hope for the best, blame the tool. In every cohort we run, the turning point is the same — the moment someone stops asking questions and starts writing briefs. Here are the five patterns that get them there.
1. Give it a role before you give it a task
“Write a follow-up email” produces something generic. “You are an account manager known for warm, low-pressure follow-ups” produces something usable. The role sets the register, the vocabulary, and the judgement calls — the things you'd otherwise spend three revision rounds fixing.
Think about what a role actually carries. An account manager doesn't just write differently from a lawyer — they weigh different risks, apologise differently, close differently. When you hand the model a role, you hand it a thousand small decisions you'd otherwise have to spell out one by one. It's the difference between hiring a specialist and micromanaging a temp.
In class we run the same task through three roles — a salesperson, a customer-success lead, and a debt collector — and read the three emails aloud. Same request, three unmistakably different letters. Nobody forgets the pattern after that.
2. Context is not optional
The model can't know your proposal went out ten days ago, that the client is worth keeping, or that the last email already said “just checking in.” Two sentences of context routinely improve output more than any clever instruction. If a new colleague would need to know it, the model needs to know it.
This is the pattern professionals resist longest, because it feels like typing. It isn't — it's triage. You already know which three facts matter: what happened, what's at stake, and what's already been tried. Give it those three and stop. Context is not a data dump; a page of history buries the signal as surely as no history at all.
If a new colleague would need to know it, the model needs to know it.
A useful test from our corporate cohorts: before you send a prompt, ask whether a sharp intern reading it cold could do the task. If they'd have to come back with questions, the model is already guessing — and it guesses confidently.
3. Specify the format, or inherit the model's
Left alone, AI produces medium-length prose with bullet points. If you need a table, a 120-word email, or a timed agenda, say so — precisely. Format instructions are the cheapest quality win in prompting: they cost one line and save one rewrite.
Be numeric where you can. “Short” means nothing; “under 120 words” is a contract. “A few options” invites three paragraphs of hedging; “exactly three options, one line each, boldest first” gets you something you can forward without editing. The model honours precise constraints with almost embarrassing obedience — it just never volunteers them.
4. Guardrails: say what it must not do
Everyone skips this section, and it's the one that separates professionals. “No fake urgency.” “Flag anything you're unsure of instead of guessing.” “No jargon.” Negative instructions catch the failure modes you already know are coming — because you've seen them in every first draft.
Your guardrails are personal, and they accumulate. One executive in our January cohort keeps a running list titled “things it keeps doing” — currently eleven lines — and pastes it into every serious prompt. Her first drafts now read like her third drafts used to. That list is worth more than any template you'll download.
Guardrails are also where responsibility lives. “Do not invent figures — mark estimates as estimates.” “If the source doesn't say it, don't claim it.” If your organisation is nervous about AI, this is the pattern that earns trust fastest.
5. Iterate in the conversation, not in your head
The first output is a first draft, not a verdict. Ask for three variations. Ask it to critique its own answer. Ask what information would make the answer better — then provide it. The professionals who get the most from AI treat it as a conversation partner with infinite patience, not a vending machine.
The habit to break is silent disappointment: reading a weak answer, closing the tab, and concluding the tool is overrated. The model can't see your frown. Tell it what fell flat — “too formal, and the second paragraph buries the ask” — and watch the next attempt. Feedback costs one sentence; starting over in Word costs your afternoon.