Last updated: 27 September 2026
Short answer: a good prompt tells the AI five things: what you want, why you want it, who it is for, what the result should look like, and what to avoid. Add an example of good output when you can, give it your source material instead of hoping it knows, and let it ask you questions before it starts on anything complicated. That one habit removes most of the back-and-forth that makes AI tools feel slow.
Why do most prompts produce generic answers?
Because the AI fills every gap you leave with the most average possible answer. "Write an email to a supplier about a late delivery" gives it no idea who the supplier is, how late the order is, whether you want to keep the relationship, or how firm to be. So it writes the email that fits everyone and suits nobody.
The model is not being lazy. It simply cannot see what is in your head. Both OpenAI's and Anthropic's own guidance start from the same point: be explicit about the goal, the context and the output you want. The rest of this guide is ways of doing that without writing an essay every time.
What should every good prompt include?
Here is a checklist you can run through in under a minute:
| Element | What it answers | Example |
|---|---|---|
| Goal | What do you want produced? | "Draft a reply to this supplier email" |
| Context | What background does the AI need? | "Order 4471 is 3 weeks late; this is the second delay this year" |
| Audience | Who will read the result? | "Their account manager, whom we want to keep working with" |
| Format | What shape should the output take? | "Under 150 words, three short paragraphs, no bullet points" |
| Constraints | What must it avoid or include? | "Ask for a firm delivery date; do not threaten to cancel" |
| Example | What does good look like? | A past email you were happy with |
You do not need all six every time. For a quick question, goal and context are enough. For anything you will send to another person, use the full list.
What does a before-and-after look like?
Before:
Write an email to a supplier about a late delivery.
After:
Draft a reply to the supplier email below. Context: order 4471 for 200 aluminium brackets is three weeks late, and this is the second delay this year. The reader is their account manager, and we want to keep the relationship. Tone: firm but polite. Ask for a confirmed delivery date by Friday and a short explanation of the cause. Do not threaten to cancel. Keep it under 150 words, in plain English.
(The supplier's original email goes here.)
The second prompt takes perhaps 40 seconds longer to write. It usually produces something you can send with light edits, instead of three rounds of "make it firmer", "shorter", "mention the order number".
How do examples change the output?
Dramatically, especially for tone and format. Describing a style in words is hard; showing it is easy. If you have a report, an email or a product description you were happy with, paste it in and say "match the tone and structure of this example."
Anthropic's documentation calls this multishot prompting and recommends wrapping each example in clear markers so the model can tell your examples from your instructions. Something as simple as this works:
Here are two examples of the style I want.
Example 1: (your first sample email or report)
Example 2: (your second sample)
Now write one for: (describe the new topic)
Use two or three varied examples rather than one. With a single example, the AI tends to copy its specific wording rather than learn the pattern. The same idea applies to image tools, where a reference image does the job of an example; see how to draw using AI tools.
How do I stop the AI from making things up?
Give it the facts instead of asking it to remember them, and give it permission to say it does not know. Three techniques help most:
- Paste the source. Instead of "What does our returns policy say about damaged goods?", paste the policy and ask the question about that text.
- Ask for quotes. "Answer using only the document below, and quote the sentence that supports each point." Unsupported claims become obvious.
- Allow uncertainty. Add "If the document does not answer this, say so." Without that line, many models will guess rather than admit a gap.
For anything factual that you will publish or act on, still check the claims yourself. These techniques reduce invented facts; they do not remove the need to verify.
How should I handle a big task?
Break it into steps, and do the steps in separate prompts. Asking for "a complete project plan with risks, budget and timeline" in one go gives a shallow version of each. Asking for them in sequence gives better results:
- "List the main phases of this project, based on the brief below."
- "For phase 2, list the tasks, dependencies and who is likely to own each."
- "Now list the top five risks for the whole plan, with a mitigation for each."
Each answer becomes context for the next prompt, and you can correct course between steps instead of discovering a wrong assumption at the end. Both providers' guides describe this approach, often called prompt chaining.
When should I let the AI ask me questions first?
Whenever the task depends on details you have not thought through yet. End your prompt with:
Before you start, ask me up to five questions that would help you do this well.
This is the single most useful line for complex requests like job descriptions, training material or a website page. The questions often surface decisions you had not made, such as who the audience really is or what the budget is, which would otherwise show up as a mediocre first draft.
How do I reuse prompts that work?
Save them. When a prompt produces a result you like, strip out the specific details and keep the structure as a template:
Draft a reply to the email below. Context: [situation]. Reader: [who, and relationship]. Tone: [tone]. The reply must: [must-haves]. Avoid: [what to avoid]. Length: [limit].
Keep templates in a note or document you can copy from. Many chat tools also let you save custom instructions or project-level context, so you do not have to repeat who you are and what your business does every time. OpenAI's developer guide makes the same point for teams: treat good prompts like any other reusable asset, and keep versions so you can see what changed.
What are the most common prompt mistakes?
| Mistake | What happens | Fix |
|---|---|---|
| No context | Generic, average answer | Add the situation and the reader |
| No format | Long, rambling output | State length and structure |
| Asking it to remember facts | Confident but wrong details | Paste the source text |
| Everything in one prompt | Shallow coverage of each part | Split into steps |
| Accepting the first draft | Mediocre result | Give specific feedback: "shorter", "less formal", "add the order number" |
For the official detail behind these patterns, Anthropic's prompting best practices and OpenAI's prompt engineering guide are both clear and worth reading.
Where do prompts go from here?
Prompts matter even more when AI tools start acting on your behalf rather than just answering. If you are curious about that shift, see what "agentic AI" really means, minus the hype. And if you are deciding where your prompts should run, on a provider's servers or your own machine, I compared the options in local vs cloud AI for a small business.
FAQ
Do prompts work the same in ChatGPT, Claude and Gemini?
The principles are the same: clear goals, context, format and examples help every model. Each provider publishes guidance with small differences, such as Anthropic suggesting XML-style tags to separate sections, but a well-structured prompt works well across all of them.
How long should a prompt be?
As long as it needs to be to remove guesswork, and no longer. A quick question can be one line. A document you will send to a client might deserve a paragraph of context plus an example.
Should I tell the AI to act as an expert?
Giving a role can help set tone and focus, for example "You are an experienced procurement manager." It works best alongside real context. A role on its own does not give the model facts it does not have.
Why does the AI give a different answer when I ask the same thing twice?
Most chat models add some randomness to their output, so wording varies between runs. If you need consistent results, use a detailed template, give examples, and ask for a fixed structure.
