How to Write a Prompt: A Step-by-Step Guide to Better ChatGPT Prompts

Here is the short version. The rest of this guide shows each step with examples:

1. Give the AI a role.
2. Add the context it can't guess.
3. Make one specific request.
4. Say what format you want.
5. Add constraints: length, tone, what to avoid.
6. Read the output, fix the weakest line, and refine.

Usually a prompt is text, but newer models also accept images and audio. It can be one line ("Summarize this in 5 bullet points") or a full brief with background, examples and rules.

Learning how to write AI prompts starts with a simple shift in thinking. Wharton professor Ethan Mollick describes working with these tools as programming with words [1]. That's the right mental model. Your words are the instructions, and the model follows them literally, including the parts you forgot to say.

Prompt engineering is the practice of designing and refining prompts so the model gives you what you need. For most people it's less "engineering" than clear briefing. You're doing what a good manager does when handing a task to a new hire.

Prompts come in a few common types:

For how these fit into named frameworks like RACE and CRISPE, see our prompt frameworks and techniques hub.

Here's the misunderstanding behind most bad prompts: an LLM is not a search engine. Google retrieves pages that already exist. A language model generates new text by predicting what's likely to come next. It's working from patterns, not looking anything up.

That has two consequences:

• The same prompt can give different answers. The output is probabilistic, not fixed.
• When your prompt leaves gaps, the model fills them with the most average answer. That's why vague prompts produce such generic text.

So a search-style query like "best social media post ideas" gets you a list anyone could have written. Add who you are, who you're writing for and what you've already tried, and the answer changes completely.

The payoff for getting this right is measurable. In a 2023 field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on suitable tasks produced results rated more than 40% higher in quality than a control group. They also finished 12.2% more tasks and worked 25.1% faster [2]. On a task outside the AI's abilities, though, AI users did worse. Knowing how to prompt is part of knowing when AI will actually help.

If you want to know how to write effective AI prompts, this is the part worth memorizing. Every prompt framework you'll read says roughly the same thing with different letters. Strip away the acronyms and good prompts share five elements:

• Role: who the AI should act as. "You're a project manager at a small web agency" sets vocabulary, tone and assumptions.
• Context: the background it can't know. Who's reading, what happened, what's already decided.
• Request: the one thing you want produced, stated plainly.
• Format: the shape of the answer. Length, sections, table columns, bullet points.
• Constraints: the guardrails. What to avoid, what to include, and what to do when it doesn't know.

Not every prompt needs all five. "Translate this into Spanish" works fine on its own. The more a task depends on judgment, audience or tone, the more of these elements you'll need. We break down the ideal order of these elements in how to structure AI prompts.

These six steps are the core of how to write ChatGPT prompts that land on the first or second try. They work the same way in Claude, Gemini or any other chat tool.

To show the difference these steps make, we ran two prompts for the same job through Claude on September 27, 2026. The task was telling a client their website launch is delayed.

The weak prompt:

Write an email to a client saying the project is delayed.

It produced exactly what you'd expect: "Subject: Project Update," a "hope this email finds you well" opener, an apology, and a promise that the team is "working diligently." It gave no new date, no reason and no next step. It's an email the client would have to reply to just to find out what's happening.

The strong prompt:

You're a project manager at a small web agency writing to a long-time client. Context: their launch moves from 3 October to 17 October because the payment integration failed security testing. They have a promo planned for 5 October. Task: write the email telling them. Format: subject line, then under 120 words in three short paragraphs. Constraints: explain why in one plain sentence, offer a temporary landing page for the promo, no "hope this finds you well," end with one clear question.

The output named the new date in the subject line and gave the reason in one sentence. It also offered a landing page so the 5 October promo could still run, and it ended with a yes/no question and a deadline. It came in at under 100 words.

Notice what changed. The model didn't get smarter between the two runs. We simply told it what we knew. That's really all there is to how to write a good prompt.

The six steps get you a solid first prompt. These best practices for writing effective AI prompts help with everything after that.

This is where most people start with AI, and where lazy prompts show up most. Ask ChatGPT to "make this better" and you'll get something smoother but blander. It's often sprinkled with words like delve and synergy that make readers suspect a machine wrote it.

The fix is to use AI as an editor, not a ghostwriter:

• Protect your voice. "Act as a line editor. Tighten this for clarity and concision without changing my voice or adding new ideas. List your changes after the rewrite."
• Show it how you write. Paste two or three samples of your own writing first and ask it to match that style.
• Ban the tells. Keep a list of overused words (delve, synergy, leverage, game-changer) and tell the model not to use them. Put the list in your custom instructions so you never retype it.
• Name the reader. "Rewrite this for first-time managers. Warm, direct, no jargon."
• Work in pieces. Edit one paragraph or section at a time. Whole-document rewrites drift.
• Ask for options, not a verdict. "Give me three alternative openings, each under 25 words."

If you'd rather not write the prompt at all, our AI paragraph rewriter is built for exactly this kind of clarity edit. For the human side of the craft, see how to improve writing flow and our AI rewriting best practices.

The method doesn't change at work. What changes is how much context you need to supply, because business output has to be accurate and on-brand.

As long as it needs to be, and no longer. There's no ideal length, only relevance:

Every extra detail competes for the model's attention, so cut anything that doesn't change the answer.

How long can ChatGPT prompts be? Far longer than you'll usually need. Current models accept very long inputs, often an entire report. Exact limits vary by model and plan, so check your tool before pasting something huge. The real question isn't whether it fits. It's whether every part of it is relevant.

A hallucination is when AI states something made up with complete confidence. It happens because generating plausible text is what these models do. Andrej Karpathy made this point in a December 2023 post on X, describing LLMs as "dream machines." In 2023, the tech site CNET had to correct dozens of AI-written finance articles after errors were found [7]. The lesson: review AI output with a critical eye, especially facts, numbers and quotes.

You can't switch hallucinations off, but you can make them much rarer:

1. Give it the source. Paste the reference text and say: "Only use the text below. If the answer isn't there, say you don't know." This is the idea behind retrieval-augmented generation (RAG) [8], and it's the single most effective fix.
2. Ask for citations. "Cite the source for each claim" makes unsupported statements easy to spot.
3. Use code for math. Ask the model to calculate with Python rather than doing arithmetic in its head.
4. Give it a way out. An instruction like "Write not in the source if you can't find it" removes the pressure to invent.
5. Ask for reasoning, not reassurance. "Walk me through how you got this" catches more errors than "Are you sure?"

For large datasets, the same logic applies. Ask ChatGPT to use its data analysis tool and work in small batches. Tell it never to truncate or use placeholder text.

Most bad outputs come from the first three rows. Fix those and you'll fix most of your results.

Good prompts for ChatGPT are worth keeping. Save your best ones, with a note on what they're for, in a doc or folder. Over time you'll have a personal library that makes your output consistent week to week. A tuned prompt beats a new one almost every time.

For rules that apply to nearly everything, use custom instructions or memory in ChatGPT (Claude and Gemini have equivalents). Your default tone, your field, your banned-word list: set them once and stop retyping them. And after a long session, ask the AI to "summarize the key decisions from this chat" before you move on. It's a handy block to paste into a fresh conversation.

If a prompt isn't working in one tool, try it in another. Models respond differently to the same wording, and the techniques in this guide transfer across all of them. Our ChatGPT vs Claude writing comparison covers where each one is strongest. For more copy-and-adapt starting points, browse our ChatGPT prompt templates.

Reading about prompting only gets you so far. The fastest way to improve your prompt engineering skills is deliberate practice on real work:

• Use real tasks, not made-up exercises. You'll notice what's weak because you care about the result.
• Get feedback. Ask a colleague, or ask the AI: "Critique this prompt. What's ambiguous?"
• Change one thing at a time, so you learn what actually made the difference.
• Keep what works in your library.

A caution from the research, too. Writing in Harvard Business Review, Oguz A. Acar argued that defining the problem clearly matters more than polishing prompt wording [9]. As AI tools get better at filling in prompts for us, that skill is what lasts. A beautifully worded prompt for the wrong problem still gets the wrong answer.

Learning how to write a prompt isn't about memorizing tricks. It's about briefing the AI the way you'd brief a sharp new colleague: who to be, what's going on, what you need, what it should look like, and what to avoid. Then look at what comes back and fix the one thing that's off.

Pick your next real task and try the six steps. That's where ChatGPT prompts for better answers start. If you want a head start, the TextToolsAI prompt generator will build the structure for you, and you can refine from there. For more guides and templates, explore our hubs.

1. Mollick, E. (2023). How to use AI to do practical stuff: A new guide. One Useful Thing.
2. Dell'Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013.
3. Brown, T., et al. (2020). Language Models are Few-Shot Learners. arXiv:2005.14165.
4. Wei, J., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903.
5. Kojima, T., et al. (2022). Large Language Models are Zero-Shot Reasoners. arXiv:2205.11916.
6. Waite, S. (2023). Confidence-based questioning prompt. LinkedIn.
7. Thorbecke, C. (2023). Plagued with errors: A news outlet's decision to write stories with AI backfires. CNN Business.
8. Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401.
9. Acar, O. A. (2023). AI Prompt Engineering Isn't the Future. Harvard Business Review.