Most people talk to ChatGPT like they're texting a psychic. They type "write me an email," hit send, and wait for it to read their mind. Then they act betrayed when it hands back something a nervous intern would be too shy to send.
Here's the thing: ChatGPT didn't fail. The prompt did. A vague ask gets a vague answer, every time. The good news: a good prompt isn't a dark art or a 40-line "mega-prompt" you buy off the internet. It's a recipe with five ingredients. Learn it once, use it forever.
The recipe: five ingredients
A strong prompt usually has five parts. You won't need all five every time, but when the output matters, the more you give, the less you'll be let down.
1. Role: who should ChatGPT be?
Tell it who to act as, and who you are. "You're my assistant, writing as me" lands very differently from "You're a sharp financial analyst." Same tool, different hat.
2. Task: what exactly do you want?
One clear job. Not "help me with this client" but "draft a two-line reply that says no without burning the relationship." Vague verbs like handle, improve, and look at are where good answers go to die.
3. Context: what does it need to know?
This is the part everyone skips, and it's the one that matters most. Who's the audience? What's the backstory? What are the constraints? ChatGPT may remember your job title, but it knows nothing about this Tuesday. Tell it, or attach the file that does.
4. Examples: show, don't just tell
If you have a sample you like (an old email, a report in the right style), paste it. "Match this tone" beats three paragraphs describing the tone.
5. Format: what should the answer look like?
Say the shape you want: length, structure, tone. "Three subject lines and a body under 120 words" gets you exactly that. "Some ideas" gets you a wall of text and a small headache.
Watch it work
Say a client was promised a report on Friday. It'll be late. You need to send the email nobody enjoys sending.
The lazy prompt:
Write an email to my client about the delay.
You'll get something. It'll open with "I hope this email finds you well," bury the point in paragraph three, and sound like a fax machine. Now the recipe version:
You're my assistant, writing as me. I'm an account manager, warm but straight-talking.
Task: draft a short email telling a client their report will be two days late.
Context: the client is Devlin & Co. We promised Friday; it'll land Tuesday. The delay is on us: a data source broke. They matter to us and they're a little anxious. I don't want to grovel, but I want them to feel in safe hands.
Format: three subject lines and a body under 120 words. Plain English, no "please be advised." End with a clear next step.
Ten seconds more typing. The second one comes back short, honest, and nearly ready to send, in a writing block you can tweak directly.
OpenAI's version: four blocks
OpenAI's own prompting guide splits a prompt into four blocks: goal, context, output, and boundaries. Same idea, different labels. Goal is our task, output is our format. The one worth borrowing is boundaries: one or two hard rules, such as "use only the figures in the attached file," "keep the prices unchanged," or "draft it, don't send it." Add a boundary when a mistake would cost you.
Four ways it goes wrong
"Make it better." Better how? Shorter, warmer, less like a robot? Name it.
Skipping the backstory. You know the client is furious and the deadline moved twice. ChatGPT doesn't.
Asking for everything at once. "Write the strategy, the email, and the deck" gets you three half-finished things. One job per prompt, then build on the result. (That's the next lesson.)
Forgetting the format. If you don't say how long or what shape, you get ChatGPT's guess. Its guess is usually "longer than you wanted."
The template
Keep this somewhere you can grab it:
Role: Who should ChatGPT be? And who are you?
Task: What exactly do you want? (one clear job)
Context: What must it know? (audience, backstory, constraints, files)
Examples: Got a sample you like? Paste it.
Format: What should the answer look like? (length, structure, tone)
Quick questions don't need all five. But the moment the output matters, run down the list. Things you'd repeat in every prompt, like your role and tone, belong in your custom instructions instead.
Try it
Grab a real email you've been putting off. Write it two ways: first one lazy line, then the recipe version with role, task, context, and format. Send both, each in a fresh chat.
Read them side by side. The gap is exactly the value of ninety seconds of thinking before you hit send.