Phase 0 · Foundations
Prompting that works
By the end, you'll be able to…
- Use a simple role-task-context-format framework to write clearer prompts
- Rewrite a vague prompt into a specific one that gets a usable answer first time
- Improve a mediocre reply by iterating instead of starting over
Why it matters
The single biggest difference between people who find AI useless and people who find it indispensable isn't the tool; it's how they ask. A vague prompt gets a vague answer. This lesson gives you a framework you can lean on every time, and three real before-and-after rewrites so you can see exactly what 'better' looks like.
Why your prompt does most of the work
A model predicts a reply from the text you give it. Give it thin, generic text and it has nothing to aim at, so it produces the safest, blandest, most average answer, which is exactly the disappointing "it just wrote waffle" result beginners complain about. Give it a rich, specific prompt and you narrow the target dramatically. You're not hoping for a good answer; you're describing one.
The good news: you don't need clever tricks or secret words. You need to say what you actually want. A simple framework makes that automatic.
The framework: role, task, context, format
Four things, in any order, cover almost everything a good prompt needs.
- Role: who you want the tool to be, or who it's writing for. "You're an experienced operations manager." "Write this for a customer who's frustrated." Setting a role shapes tone and judgement.
- Task: the actual thing to do, stated as a clear instruction. Not "the sales figures" but "find the three biggest month-on-month drops and explain each in one sentence."
- Context: the background and material the tool needs: the situation, the audience, any text or data to work from, the constraints. This is where most weak prompts fall down; they leave the tool guessing.
- Format: what the output should look like. Bullets or prose? How long? A table? A subject line and three paragraphs? Say so, and you'll stop getting a wall of text when you wanted a list.
You won't always need all four, and you needn't label them. But when an answer disappoints, run down the list; you'll almost always find the missing one.
Two more habits multiply the framework's power:
- Be specific. Numbers, names, limits, examples. "About 100 words" beats "short". "For a non-technical colleague" beats "simple".
- Show an example. If you want output in a particular shape or voice, paste a small example of it. Models copy patterns brilliantly; one good example is worth a paragraph of description.
Rewrite 1: the vague request
The most common mistake: asking for a thing with none of the detail that would make it good.
Write something about our sale.
Why this works: 'Write something about our sale' gives the model no audience, no facts, no length and no format, so it can only produce generic filler.
The reply is exactly as empty as the request: a few paragraphs of "Don't miss our amazing sale!" that could belong to any shop on earth. Now the same intention, with role, task, context and format supplied:
You're writing a short marketing email for Fernway's existing customers, friendly, not pushy. We're running 20% off all annual subscriptions until the end of August. Write a subject line and about 90 words of body copy. End with a single clear call to action to renew early.
Why this works: A named audience, the concrete facts, a length and a format turn 'write something' into a describable target the model can actually hit.
Now you get a usable draft: a real subject line, the right facts, the right length, a clear ask. The difference isn't a cleverer tool; it's that the second prompt described the answer.
Rewrite 2: the context-free question
Sometimes the task is clear but the tool has nothing to work from, so it invents plausibly.
Write a reply to an unhappy customer.
Why this works: Asking for a reply with none of the actual situation forces the model to guess who the customer is and what happened, so it writes a generic template.
You'll get a one-size-fits-nobody template full of "[insert issue here]" gaps. Feed it the real situation instead, stripped of anything confidential, and it can write something you'd actually send:
A customer emailed to say they logged a support issue two weeks ago and never heard back. They're understandably annoyed but polite. I want to apologise sincerely without making excuses, explain we're fixing how we track issues, and offer to sort their problem this week. Write a warm, professional reply of about 120 words. Don't invent details I haven't given you.
Why this works: Giving the specific situation, the tone you want and the outcome you can offer lets the model write a genuine reply instead of a fill-in-the-blanks shell.
Notice the last line, don't invent details. It's a small guardrail that stops the tool from confidently filling gaps with things that never happened.
Rewrite 3: the format-free dump
Here the task and context are fine, but you'll drown in prose when you wanted something scannable.
Tell me about these meeting notes. [paste]
Why this works: 'Tell me about these notes' invites a rambling paragraph; without a shape, the model defaults to prose you then have to reorganise yourself.
The reply is a dense summary you still have to pull apart. Ask for the exact shape you need and the tool does that work for you:
From the meeting notes below, give me: (1) a three-sentence summary a manager could read at a glance, then (2) a table of every action with three columns: Action, Owner, Due date. Mark any action with no owner as 'UNASSIGNED'. Notes: [paste].
Why this works: Naming the structure (a short summary, then a table of actions with owners and dates) gives you output you can use as-is instead of reformatting by hand.
Iterate, don't start over
Your first answer is a draft, not a verdict. When it's close but not right, the fastest fix is almost never a brand-new prompt or a different tool; it's a short follow-up in the same chat, where the tool can still see everything above.
Good iteration sounds like plain feedback: "Shorter, half the length." "Warmer, less formal." "You missed the August deadline; add it." "Turn the third point into a table." Each nudge steers the existing draft closer, and because the earlier context is still on the desk, you keep everything that was already working.
Good start. Two changes: cut it to about 70 words, and make the tone a little warmer, less corporate. Keep the call to action.
Why this works: Concrete, specific feedback on the reply you already have is faster and better than re-explaining from scratch; the model keeps the good parts and adjusts only what you flagged.
Try it now
Common mistakes
- Blaming the tool for a vague prompt. A generic answer is usually a generic request wearing a disguise. Run the four-part check before deciding the tool "can't do it".
- Retyping the same prompt hoping for better. Same input, same kind of output. Change the prompt or give feedback; don't just resend.
- Abandoning a near-miss. A close draft is a follow-up away from right. Starting a fresh chat throws away context the tool could have built on.
- Over-trusting a polished first draft. A specific prompt gets you a fluent, confident answer, but fluent isn't the same as correct, especially for facts and figures. Read it critically and check anything load-bearing before you send it on; a great prompt improves the writing, not the truthfulness.
Keeping current
The framework is durable: role, task, context and format will still work in years. What changes is what tools can take as input (voice, files, live data) and the odd model-specific quirk. When you want the latest on getting good results from a specific tool, check its official prompting guidance, such as OpenAI's prompt engineering guide. Accurate as of 13 July 2026.