A good prompt has no magic formula. What changes the answer is what you put in: your goal, your context, and the shape of the reply.

You open ChatGPT or Claude, and you might type:

write a client delay email

You get back a polite, generic, slightly hollow paragraph. You spend ten minutes reshaping it to fit your actual client. Frustrating.

Try this instead:

10|> Write an email to a client announcing a two-week delay on their project.

Context: small construction company in Quebec. Client: Mr. Tremblay, long-time repeat customer. Cause: a delay from our steel supplier. Respectful and honest tone, without over-apologizing, no new commitment date until the supplier has confirmed one. Length: 150 words max, in Canadian English. Signature: Alain, project lead.

The email that comes back sounds right, in your voice, ready to send after light edits. The prompt did almost all the work.

What actually matters

You've probably heard the "You are an expert in…" formula. It's optional, and often unnecessary on 2026-era models. What actually moves the answer, in order:

20|1. **A clear goal.** A precise verb and a precise deliverable. "Write an email" beats "help me with an email".

2. Context the model can't guess. Who, when, where, why. This is where you gain 80% of the quality. 3. Shape constraints. Length, format (bullets, table, email), audience, language, tone, what to avoid. 4. Raw material. Paste in the facts, snippets or notes you already have. The model works far better with data than with your memories. 5. An uncertainty instruction. For anything factual, add: "If you aren't sure, say so instead of making something up."

A role sentence ("You're a copywriter for…") doesn't hurt, but it never fixes thin context.

Three reusable examples

Meeting summary.

30|> Summarize the transcript below as:
  1. 5 key points
  2. Decisions made (with owner)
  3. Open action items (with owner and due date if mentioned) Audience: management team. Short bullets. If a decision is unclear, mark it "to confirm" rather than guessing.

Reply to a vendor.

Draft a reply to this email proposing an 8% price increase. Our margin can absorb 3% max. Diplomatic tone, open to negotiation, no numeric commitment until we've talked internally.

Faithful translation.

40|> Translate the text below from French to Canadian English for a business audience. Keep meaning and tone, don't add, don't cut. If a passage is ambiguous, add [to verify] next to it.

Each prompt lives in a goal, some context, a shape, and an uncertainty rule when it matters.

The real move: iterate

A well-built first prompt gets you a usable v1. The real gain comes from the second turn: "shorter", "more direct", "replace line 2 with…", "redo assuming the client pushes back on price". Two or three passes and you have a result that sounds like you.

What not to do

  • Paste confidential data (client, staff, contract) into a free account whose data policy you haven't read.

  • Ask for numbers without giving a source: the model will invent rather than admit it doesn't know.

  • Trust a quote, a law, or a statistic without verifying it elsewhere.

    50|## Try this this week

Pick a repetitive task from your week (template email, summary, plan). Write the prompt with the five elements above. Test it, adjust, and save the version that works in a "my prompts" file. Five saved prompts is already a personal library that pays you back every day.

More on this

Context quality drives answer quality: why context in your prompt is crucial, with before/after examples. To go further, six prompting techniques for everyday work. Natural next step: have AI summarize a PDF (and check it).