How do I write better prompts for ChatGPT?

Short answer: give it context, be specific about what you want, and treat the first answer as a draft to iterate on. Most bad results come from vague prompts, not from a weak model. A few sentences of background and a clear description of the output you want will improve your results more than any exotic prompting trick.

Prompting is a skill, but it's not a mysterious one. It's closer to giving good instructions to a capable assistant: the clearer you are about the situation, the goal, and the constraints, the better the work comes back.

Start with context, not just the task

The single biggest upgrade most people can make is adding background before the request. "Write an email to my client" is a task with no context. "Write an email to my client explaining that the project will be two weeks late because our supplier went bankrupt, keeping the tone professional and preserving the relationship" is a task with context — and it will produce something dramatically better.

The model doesn't know your situation unless you tell it. It doesn't know who the audience is, what tone fits, what constraints matter, or what you've already tried. Every relevant detail you add narrows the space of possible answers toward the one you actually need.

A useful habit: before typing the request, ask yourself what you'd tell a smart colleague who'd been out of the loop. That briefing — the situation, the goal, the audience, the constraints — is your prompt.

Be specific about the output

Vague requests get vague answers. "Tell me about marketing" gets you a generic essay. "Give me five email subject lines for a product launch targeting small bakery owners, each under 50 characters, with a warm tone" gets you something usable.

Specify the format when it matters. If you want a table, say so. If you want bullet points instead of paragraphs, say so. If you want a specific length — "in about 200 words," "as a three-paragraph summary" — say so. The model is happy to shape its output to your needs, but it won't guess them.

Specify the audience too. "Explain this for a beginner" and "explain this for a fellow engineer" produce very different answers, and both are better than an answer aimed at nobody in particular. The audience determines the vocabulary, the depth, and which details matter.

Give it a role or perspective

One of the simplest effective techniques is assigning a point of view: "Act as a hiring manager reviewing this resume," or "You're an experienced travel planner — help me plan three days in Kyoto." This works because it activates a relevant frame — the model draws on patterns associated with that role, including its priorities and vocabulary.

Don't overthink the role assignment. It doesn't need to be elaborate. A short phrase is enough to steer the response. And you can combine it with everything else: role, context, task, format, constraints, all in one prompt.

This is also useful for getting pushback. "Act as a skeptical investor and poke holes in this business idea" will get you a more useful critique than "what do you think of my business idea," because you've explicitly licensed the model to be critical instead of encouraging.

Show, don't just tell

When you want a particular style or structure, examples beat descriptions. Instead of explaining at length what tone you want, paste a paragraph you like and say "write in this style." Instead of describing the format, show a short sample of the output format and ask it to follow the pattern.

This is sometimes called few-shot prompting, but you don't need the jargon. The intuition is ordinary: it's easier to match an example than to interpret an abstract description. Humans work this way too — "make it look like this" is clearer than a paragraph of design specs.

Even one example transforms the output. If you're generating product descriptions, show one good one. If you're drafting social posts, show the style you want. The model is excellent at pattern-matching; give it a pattern.

Iterate instead of starting over

Most people treat prompting as one-shot: they write a prompt, read the answer, and either accept it or give up. The better approach is conversational. The first answer is a draft. Respond to it the way you'd respond to a colleague's draft: "that's close, but make it shorter," "good, now add a section on pricing," "the tone is too formal — loosen it up."

Each round of feedback narrows in on what you want, and it's usually faster than trying to write the perfect prompt on the first attempt. In fact, trying to perfect the initial prompt is often counterproductive — you spend ten minutes engineering a prompt when two quick iterations would have gotten you there.

This also means you should read the answer before judging the prompt. Sometimes the prompt was fine and the answer just needs a nudge. Iteration is the normal workflow, not a sign of failure.

Break big tasks into steps

For complex work, don't ask for everything at once. If you want a business plan, start with the outline, refine it, then ask for each section. If you want code, describe the overall structure first, then build it piece by piece.

There are two reasons this works better. First, each step gets the model's full attention instead of splitting it across a huge task. Second, you can catch misunderstandings early — if the outline is wrong, you fix it before the model writes two thousand words in the wrong direction.

You can also ask the model to reason through something before answering: "think through the pros and cons first, then give me your recommendation." This tends to produce more considered answers for judgment-heavy questions, because the model works through the analysis instead of jumping to a conclusion.

A simple template you can reuse

If you want a structure to fall back on, here's one that covers the essentials without becoming a chore to fill in:

1. Context: what's the situation? (A sentence or two of background.)
2. Task: what do you want produced?
3. Audience: who is it for?
4. Format and length: what should it look like, and how long?
5. Constraints: tone, things to avoid, things to include.
6. Example (optional): here's what good looks like.

In practice it reads like this: "I'm a freelance designer writing a proposal for a restaurant rebrand (context). Draft the project timeline section (task) for the restaurant owner, who isn't technical (audience). Keep it to five or six bullet points, plain language, no jargon (format and constraints)."

You won't use all six elements every time — quick questions don't need them. But when a result matters and the first attempt was weak, running through this checklist almost always reveals what was missing. Nine times out of ten, it's the context or the audience.

The template also helps you avoid the most common prompting mistakes. Beyond vagueness, a few specific habits produce bad results. One is asking multiple unrelated questions in a single prompt — the answers come back shallow because attention is split. Ask one thing, get a good answer, then ask the next.

Another is overloading the prompt with contradictory instructions: "be thorough but brief, formal but friendly, detailed but high-level." The model tries to satisfy everything and satisfies nothing. Pick the two or three constraints that actually matter and drop the rest.

A subtler one is leading the witness. "Don't you agree that remote work is always better?" will get you agreement, not analysis. If you want a real answer, ask neutrally: "what are the strongest arguments for and against remote work for a ten-person team?" You'll get a better answer, and you might learn something.

Finally, people often forget to say what the output is for. "Summarize this report" produces a different summary than "summarize this report so I can decide whether to invest." The purpose shapes which details matter. Include it.

Know what it's bad at

Good prompting includes knowing when not to bother. The model can confidently produce wrong facts, fake citations, and plausible-sounding nonsense — especially about niche topics, recent events, or anything requiring precise data. Always verify factual claims that matter, and never trust a citation you haven't checked.

It's also bad at knowing what it doesn't know. It won't reliably tell you when it's guessing. So for anything where accuracy matters — medical, legal, financial, or technical details — treat the output as a starting point for your own verification, not as an answer.

And it's bad at true originality. It remixes patterns from its training data brilliantly, but it won't have a genuinely novel insight or know your unique situation better than you do. The best use is as a thought partner and draft-writer, not as an oracle. Your judgment stays in the loop — that's the deal.

Better prompts won't fix a bad question, and no prompt fixes over-trust. But clear context, specific instructions, a useful perspective, an example or two, and a willingness to iterate will get you dramatically better results than most people get. It's not magic. It's just communication, done carefully.