A prompt is the instruction you give an AI. When you state what you want to make, which materials to use, who will read it, what form it should take, and what it must respect, the answer becomes far more accurate, and it becomes clearer how to check the answer. This chapter compares a weak prompt with a stronger one, then walks through a meeting-notes summary, from writing the instruction to checking the response.
Meeting key points, an email draft, a comparison table, and so on.
The facts that are settled, and what is still unknown.
Whether the answer matches the materials, and whether anything is missing or breaks a condition.
What a prompt can and cannot change
"Summarize the meeting" doesn't tell the AI who will read it, whether you want the decisions, or whether you want the discussion recorded in detail. "Split the summary for people who missed the meeting into decisions, open items, and next steps" states the result you want.
The official documentation from OpenAI and Anthropic also describes methods such as giving clear instructions, providing relevant materials, specifying the output format, and including examples. However, a more detailed prompt does not guarantee a correct answer, and the same prompt can produce slightly different answers each time.[1][3]
- Who the output is for and what to make
- The materials to reference and the form of the answer
- The conditions to follow and the items to check
- Whether facts not in the materials, or latest information, are correct
- Professional qualifications or hands-on experience
- Available features, permissions, and usage limits
For example, writing "Answer as an experienced lawyer" does not give the AI a legal license. Likewise, a prompt cannot unlock a search feature that isn't available to you or let the AI access files you aren't allowed to use. To learn how to choose a tool, see How to choose AI tools.
The five elements of a prompt
These five elements are a way to organize your instructions. You don't have to fill in all of them every time. Start with the goal and materials, then add the conditions you need.
Decide what the finished output will be, such as a summary, a draft, or a comparison.
Example: Summarize the key points of the meeting for people who missed it.
Separate the facts you can use from the points that are not yet decided. Set reference materials apart from the instructions.
Example: The budget has not been approved yet. Do not add information that is not in the notes.
Tell the AI what the reader already knows and what viewpoint to take. Assigning a role also goes here.
Example: Use wording that colleagues who did not attend can understand.
Choose a format that suits the situation, such as headings, bullet points, or a table.
Example: Three headings: decisions, open items, and next steps.
State what is required, what to avoid, and how to prioritize when conditions conflict.
Example: Do not promise a delivery date that hasn't been decided. Record it as an item to confirm.
Rather than giving all the background, choose the materials that this task needs. Whether you may enter real confidential information is a separate decision from how you write the prompt. Before you use work documents, also read Checking what information you enter and privacy.
Seven techniques to use
Writing a long prompt is not the goal in itself. Choose the technique that fits the problem you have, and then check the answer.
1 Describe the finished result specifically
Instead of "Give me some good ideas," say what the ideas are for and what you will look at when you choose one.
Please write an announcement for first-time participants of a book club. The meeting date is not decided yet. Do not invent a date or time. Separate how to join from the items to confirm before publishing.
Here, the prompt specifies what to make and the conditions to respect.
2 Separate materials from instructions
Use headings or dividers to separate "Task," "Materials," and "Output format." OpenAI describes separating content with Markdown or XML, and Anthropic describes using XML tags. Beginners can simply start with bracketed labels such as [Task].[1][3]
[Task] Summarize the key points using only the materials below.
[Materials] Paste here the material you have confirmed can be entered.
[Output format] Separate decided items from open items.
Note that even with separators, the AI may follow an instruction hidden inside the materials. Be careful when you paste text whose source you don't know.
3 Specify the reader and point of view
Tell the AI how difficult the explanation should be or what to look for, such as "for beginners" or "point out gaps from an editor's point of view." Anthropic describes assigning a role as a way to shape the AI's behavior and writing style.[3]
Explain it as if to someone attending a meeting for the first time. Briefly explain any technical terms. Do not leave out the conditions needed for a decision.
Assigning a role does not add expertise, so check specialized conclusions against the original materials or with the person in charge.
4 Show correct input-output examples when needed
When a format or classification standard is hard to explain in words, add examples of the input and the output you expect. Giving several examples is called few-shot prompting. Choose examples that match the actual task and do not contradict your instructions.[1][2][3]
| Input example | Classification shown | Reason |
|---|---|---|
| The venue has been decided as Room A. | Decision | It is explicitly stated as decided. |
| Room A is being considered as a candidate venue. | Open item | It is a candidate, not a decision. |
| The venue is decided, but the start time is not. | Split by item | The venue is decided; the start time is open. |
Be careful not to give an incorrect classification as an example. There is no fixed number of examples. You can also try without examples first and add one only for a standard that doesn't come across in words.
5 Split complex work into checkable steps
For example, if you split the work into "extract the conditions from the materials, check the conditions, then write a draft," you can see where an error entered. If an intermediate judgment matters, have a person check it before moving on.
First, extract the decided items and open items from this material. Do not write the announcement yet. After I review the extracted items, write the announcement.
Many recent AI models have a built-in way of reasoning before they answer, so OpenAI's documentation says you don't need to ask them to "think step by step."[2] Instead, asking for the parts you need to verify, such as the sections of the materials used, the calculation formula and units, and the assumptions made, is useful.
6 Specify what to check when asking for a review
Instead of only asking "Is anything wrong?", specify what to compare it against.
Compare the draft with the original notes. Check whether any undecided date is presented as confirmed, and whether any budget that has not been approved is presented as approved. Point out the sentences in question and the matching parts of the notes.
Even if the AI says "No problems," a person should check important figures, dates, and commitments against the original materials.
7 Ask AI to help write the prompt
Tell the AI what you want to make, and ask it to question you about any missing conditions. Reread the suggested prompt to make sure it doesn't add facts you don't know or unnecessary conditions.
Help me write a prompt for creating an event announcement. I want to organize the purpose, the readers, the facts that are decided, the points that are not decided, and the output format. Ask me about any missing information, and do not guess dates or amounts.
Questions you can't answer yourself can stay open for now.
Before / After — the same request can change this much
Explanations of techniques can be hard to picture, so compare a request before and after revision. On the left is a vague request, and on the right is one with conditions added. (Names and numbers are examples.)
Example 1: Writing an apology email
| Before | After |
|---|---|
| "Write an apology email." → It is unclear who is being apologized to or for what, so the result is a generic message that could fit anywhere. |
"Write an apology email to Mr. Tanaka, the client's department head, for the delivery being 3 days late. The cause was a delay in logistics, and the delivery is expected next Monday. Do not include any promises or compensation that I have not written." → You get an apology draft that fits the situation. Before sending, check the dates and promises against your records. |
Example 2: Brainstorming a project idea
| Before | After |
|---|---|
| "Give me ideas for a new product." → It is unclear what industry or product this is for, so the suggestions stay vague. |
"Propose 5 subscription health food products for women in their 20s and 30s, priced at 3,000 yen per month or less. For each one, add a short note on whose problem it solves and what that problem is." → You get focused ideas that can serve as a starting point for a planning meeting. |
Example 3: Learning and research
| Before | After |
|---|---|
| "Teach me machine learning." → The answer tends to be dense and full of technical terms. |
"Explain how machine learning works so that a middle school student could understand it, using a recipe as an analogy. Do not use technical terms, and include two concrete examples." → You get an explanation that is easy to read right away, using a familiar comparison. |
What these examples have in common is that they make clear who it is for, what to make, how to make it, and in what form. Next, let's practice this approach by summarizing meeting notes.
Practice: Summarize meeting notes for people who missed the meeting
- The venue for the book club has been decided as Room A.
- The start time is not yet decided. The person in charge will present options at next week's meeting.
- The printing cost for materials is at the proposal stage; neither the amount nor approval has been decided.
- We want to tell people who missed the meeting what has been decided and what is still open.
Summarize the meeting briefly.
It doesn't say what to keep, which materials to use, or how to handle the points that are not decided.
Using only the notes below, summarize the key points of the meeting for people who missed it. Use bullet points under three headings: decisions, open items, and next steps. Do not invent dates or amounts; if something is not in the notes, mark it as "not yet decided."
[Notes]
(Paste the practice notes above here.)
It includes the conditions you need to compare the answer against the original notes.
Sample answer
- The venue for the book club is Room A.
- Start time; printing cost amount and approval.
- The person in charge will present start time options at next week's meeting.
- Printing cost: confirm the amount and when approval will happen with the person in charge.
When you try this, paste the prompt and the notes into a new chat. The answer does not have to use the same wording as this sample. Check the content against these three points:
- Does it separate decisions from open items? Room A is decided; the start time and printing cost are open.
- Does it avoid inventing dates or amounts? The notes don't give a specific date for next week's meeting either.
- Does it avoid turning a plan into a commitment? Presenting start time options is a plan, not a confirmed event time.
If you turn this into an email, remember that having a draft and being ready to send it are different things. Drafting emails and chat messages also covers how to check what you are promising to the other person.
How to fix common problems
| Problem | What to say next or check |
|---|---|
| The AI doesn't understand what "good" means for you | Make the reader, purpose, and items to keep specific, for example "separate decisions from open items." |
| Word count and required items conflict | Decide what matters most. If everything is required, revisit the length. Also check what the AI left out. |
| The request mixes unrelated tasks | Split it by purpose. Related tasks can be combined, but structure them so that you can check each intermediate judgment. |
| You trust an answer because it cites a source | Open the original source and check that it actually says what the answer claims. A URL alone is not proof. |
| Facts not in the materials get added | Limit the materials to use, and tell the AI to mark anything unknown as "not yet decided." Compare the revised output with the materials again. |
| Instructions drift further apart the longer the conversation goes | Restate the current goal, the conditions you've adopted, and the open points. Don't carry over outdated conditions. |
Long conversations and large amounts of material may also run into input and conversation limits that depend on the service or model. Writing "Remember everything" does not remove those limits. For how to organize all the information you provide, including reference materials, past exchanges, and tool results, see Context engineering.
Check the answer and fix only what needs it
What matters most is stating specifically what was wrong, rather than simply trying more times. Sometimes you get a usable answer on the first try, and sometimes it takes several rounds of revision.
Provide the goal, materials, and conditions.
Check the output against the original materials and the conditions for the finished result.
Point out where it went wrong and what change you need.
Check the revised version again and decide whether to use it.
A correction request you can use as is
The start time is still undecided, but the answer says 2:00 p.m. Remove this time and move it back to open items. Keep the point that the venue has been decided as Room A. After the revision, check whether any dates, amounts, or approvals that are not in the original notes have been added.
If the output still doesn't match the materials after revising, try a different approach, such as shortening the materials or splitting the task.
- It is clear what to make and which materials to use.
- You have not let the AI fill in unknown points with guesses.
- The required conditions and their priorities do not contradict each other.
- You have compared dates, figures, and commitments against the original materials.
- You have confirmed what information you are sending outside and whether you are allowed to use the output.
Official references and next steps
The following are official documents from each company. They include explanations for developers (API users).
- OpenAI: Prompt engineering — structuring instructions, examples, relevant materials, and differences between models.
- OpenAI: Best practices for reasoning models — keeping instructions simple, setting clear goals, and aligning examples with instructions.
- Anthropic: Prompting best practices for Claude — guidance on clarity, separating materials, roles, examples, and thinking behavior in different models.
For task-specific prompts, see Practical prompt engineering guide. In the next chapter, you'll learn about ways to use AI at work, for study, and for creative projects.