How to Write Prompts for ChatGPT

Writing effective ChatGPT prompts comes down to being specific about the task, providing real context, and stating the format you want back — vague requests get vague answers, while specific ones get usable output on the first try. The gap between a mediocre ChatGPT user and a genuinely effective one is almost entirely in these habits, not in knowing some hidden set of tricks.

The three things every good ChatGPT prompt includes

Task, context, and format. The task is what you actually want done — stated as an action, not a topic. Context is the background ChatGPT can’t infer on its own: your audience, your constraints, what you’ve already tried, why the request matters. Format is how you want the answer structured — a list, a table, a specific length, a particular tone.

Missing any one of these usually means a follow-up message to fix what came back. Missing task specificity gets you something generically responsive but not actually useful. Missing context gets you something that’s specifically shaped but doesn’t fit your real situation. Missing format gets you the right content in the wrong shape, which often means manually restructuring output that ChatGPT could have given you correctly the first time if asked.

Why “act as an expert” prompts underdeliver

Role-playing prompts — “act as a marketing expert,” “you are a senior software engineer” — get recommended constantly in prompt engineering content, but a role alone doesn’t add information ChatGPT didn’t already have access to. The model isn’t unlocking hidden expertise by being told to roleplay; it’s adjusting tone and framing based on what “marketing expert” typically sounds like in its training data.

What actually improves output is combining a role with real specifics: “act as a marketing expert reviewing this specific campaign for a B2B SaaS audience with a $50K quarterly budget” does genuine work, because the specifics — audience, budget, campaign type — are information the model needed anyway. “Act as a marketing expert” alone mostly just changes the tone of a generic answer, giving you something that sounds more authoritative without necessarily being more useful.

This overlaps with how the system prompt works, though Custom Instructions and in-message roles operate differently. This doesn’t mean roles are useless — they can genuinely help calibrate tone and the assumed level of background knowledge in a response. It means a role is a supplement to specificity, not a substitute for it.

Iterating beats getting it perfect on the first try

A rough first prompt followed by two specific refinement messages usually beats spending five minutes crafting the “perfect” single prompt. Start with task and context, see what comes back, then give concrete feedback on what to change — this conversational approach uses ChatGPT’s memory of the thread instead of re-explaining everything each time.

The refinement itself needs to be specific to work well. “Make it better” gives the model nothing concrete to act on. “Make the second paragraph more concise and cut the closing sentence entirely” gives it exactly what changed and what to do about it. Treat each refinement message the way you’d give feedback on a real draft — specific, actionable, pointed at one or two things rather than a vague overall reaction.

How ChatGPT’s memory within a conversation actually works

Within a single conversation, ChatGPT retains everything said earlier in that thread, which is what makes iterative refinement effective — you don’t need to restate context you already gave it three messages ago. This also means earlier mistakes or unclear phrasing in your own messages can compound; if your third message assumes something you never actually said, the model may have filled that gap with an incorrect assumption that then shapes everything after it.

Across separate conversations, though, there’s no shared memory by default (Custom Instructions and any persistent memory features are the exception, and those need to be explicitly configured, not assumed). Starting a fresh chat for a genuinely new task, rather than continuing an old thread that’s drifted far from its original topic, tends to produce cleaner results than trying to redirect a long, meandering conversation.

Formatting requests that actually work

Being explicit about format upfront — “as a bulleted list,” “in a table with columns for X and Y,” “under 150 words,” “no headers, just flowing paragraphs” — saves a follow-up message almost every time. ChatGPT defaults to a reasonable-looking format based on the content of the request, but “reasonable-looking” and “what you actually needed” aren’t always the same thing.

This matters more for content going somewhere specific — a slide, an email, a spreadsheet cell — where the wrong format means real rework, not just a stylistic mismatch. If you know the destination for the output, naming it directly (“this needs to fit in a single email paragraph” or “this is going into a spreadsheet cell, so no line breaks”) often gets a more useful result than describing the abstract format you want.

Common mistakes even experienced users make

Stacking too many distinct requests into one prompt — asking for a summary, a critique, and a rewrite all in one message — tends to produce a response that does all three shallowly rather than one of them well. Breaking multi-part requests into sequential messages, even within the same conversation, usually gets better results at each individual step.

Another common mistake is assuming ChatGPT knows something it genuinely doesn’t — your company’s internal terminology, a document you haven’t actually pasted in, a previous conversation from a different chat window. If a response seems to be working from wrong assumptions, the fix is almost always adding the missing context explicitly, not rephrasing the request in a different way.

Using Custom Instructions for recurring work

If you find yourself restating the same context or preferences at the start of most conversations — your job, your writing style preferences, the fact that you always want concise answers without caveats — Custom Instructions in ChatGPT’s settings lets you set that once and have it applied automatically to every new chat. This is genuinely underused; a lot of people who use ChatGPT daily never configure this, and end up re-explaining the same background information dozens of times over.

The tradeoff is that Custom Instructions apply globally, so overly specific instructions meant for one type of task can actively hurt unrelated requests. A good practice is keeping Custom Instructions to genuinely stable preferences — tone, format defaults, your general context — and reserving task-specific detail for the actual prompt each time.

When a longer, more structured prompt is worth the extra effort

Most everyday requests genuinely don’t need an elaborate prompt structure — task, context, format in a couple of sentences is enough. But for complex, high-stakes, or one-shot requests where you won’t get a chance to iterate (a final version of something going out the door, a request where a wrong assumption would be costly to catch late), it’s worth investing more upfront: explicitly listing constraints, providing an example of the output style you want, and stating what to avoid.

The signal for when to do this isn’t complexity of the topic — it’s how costly a wrong first attempt would be. A quick brainstorm doesn’t need this treatment even if the topic is complex; a client-facing document does, even if the topic is simple.

Providing examples to anchor tone and format

When a specific style or format genuinely matters and is hard to describe precisely in words, providing one concrete example of what “good” looks like usually outperforms several sentences trying to describe it abstractly. This is especially true for tone — “professional but warm” means different things to different people and different models, while an actual example sentence in that tone removes the ambiguity entirely.

A quick way to see this structure in action is the AI prompt generator, which builds a task-context-format prompt for you as a starting point. This works well combined with the task-context-format structure: state the task and context as usual, then add “here’s an example of the tone I’m going for: [example]” before the actual request. The example does work that description alone often can’t.

Prompting differences between GPT-4-class models and lighter/faster ones

ChatGPT offers access to multiple underlying models, and prompt behavior varies between them in ways worth knowing about. Faster, lighter models tend to take instructions more literally and are more likely to miss implied context that a more capable model would infer correctly — which means being slightly more explicit and less reliant on the model “figuring out” what you meant pays off more with these models than with the most capable option.

For complex, multi-step reasoning tasks, the more capable models are worth the extra time even though they respond more slowly. For quick, well-defined tasks — reformatting text, a short factual question, a simple rewrite — the speed difference often isn’t worth the wait, and a lighter model handles these reliably. Matching the model to the actual complexity of the task, rather than always defaulting to the most capable option, is a genuinely underused habit — see the beginner prompt engineering guide if you’re still building these habits from scratch.

Try it yourself

Help me [specific task]. Context: [relevant background ChatGPT can’t infer]. Format needed: [format, e.g. “a bulleted list” or “under 200 words”].

That’s a good start, but [specific change needed]. Keep [what’s already working] the same.

FAQ

For platform-specific phrasing, see how to write image-generation prompts or browse the full prompt library.