“Write an announcement,” “Make this document look good”—short requests are easy, but because they don’t convey who the audience is or what the purpose is, the answers can often miss the mark.
In such cases, instead of having the AI write immediately, there is a way to have it identify missing information first. However, if there are too many questions, the effort required to answer them increases. The key is to narrow it down to only the questions that significantly impact the results.
■ What to convey first
Include the following three points in your request:
・What you want to complete
・Materials or conditions you currently know
・Ask questions if there is anything that needs to be confirmed before answering
Furthermore, if you tell it to “clearly state any assumptions for minor points of uncertainty,” you can reduce the back-and-forth of confirmations. This approach distinguishes between asking questions when important conditions are missing and proceeding with the work by stating assumptions when the impact is small.
■ Clarification prompt you can copy and use
Before executing the following request, please check if there is any missing information that would significantly affect the quality of the output.
Request: [What you want done]
Purpose/Usage: [What you know]
Materials/Constraints: [What you know]
Procedure:
-
If there are points of uncertainty that would significantly change the result, ask a maximum of 3 questions in order of priority.
-
For minor points of uncertainty, proceed with the work by stating a reasonable assumption in one sentence.
-
Do not fabricate facts. Mark information that cannot be confirmed as “unconfirmed.”
-
If no questions are necessary, proceed to create the deliverable as is.
When asking questions, please format them so they are easy to answer all at once.
■ Example of use: Event announcement
If the initial request is just “Write an event announcement,” neither the date nor the target audience is known. By entering “Cooking class for local beginners. To be used on SNS and store flyers. Date is undecided” into the template, the AI can confirm important items such as the date and application method.
When you receive questions, answer only what you know. If you reply to items that haven’t been decided yet with “Undecided. Please leave the section blank instead of using a placeholder,” you will get a draft that does not contain fictitious information.
■ Additional instructions for when you want to reduce questions
“Only ask questions about points of uncertainty that require stopping the work. For everything else, clearly state your assumptions and create the first draft.”
This can be used if you are asked questions every time and the work stops. Conversely, for tasks where mistakes have a large impact, such as contracts, payments, or external transmissions, use the following instruction.
If there are any unclear points regarding the amount, deadline, recipient, or content to be sent externally, do not guess; please be sure to confirm with me before creating a draft.
■ Do not blindly trust AI questions; judge their necessity
You do not need to answer every question the AI asks. Consider whether the result would change without that information. If it is not important, you can reply, “Please proceed with general assumptions and clearly state those assumptions.”
Also, to check if the AI has correctly understood a vague request, it is easier to verify if you have it summarize the request in one sentence before it begins the work.
Before you start creating, please rephrase the purpose of the request and the final deliverable in one or two sentences. If there are any differences from my request, please confirm them at that point.
■ Summary
The method of having the AI ask questions is useful for those who struggle with organizing long requirements. Limit questions to a maximum of three, focusing only on critical missing information. For minor uncertainties, proceed by explicitly stating your assumptions. Even just doing this can help reduce back-and-forth confirmation and major rework.
However, for judgments involving medicine, law, money, contracts, or documents actually sent to people, the AI’s questions alone cannot guarantee safety. Please perform final judgments and fact-checking yourself.
Reference information:
OpenAI Help Center “Prompt engineering best practices for ChatGPT”
https://help.openai.com/en/articles/10032626-prompt-engineering-best-practices-for-chatgpt
Anthropic Docs “Prompting best practices”
https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/prompt-templates-and-variables
