Blog · ChatGPT · Published

How to Write ChatGPT Prompts: Keep the Goal, Material, Requirements, and Tone Separate

According to official prompting guidance, a prompt should at minimum state what needs to be done, which materials to rely on, what form the result should take, and what tone to use. When an answer is unsatisfactory, adjust it item by item against these categories rather than rewriting the whole thing or judging the prompt on a single result.

When writing a ChatGPT prompt (the text you send to the model), first separate four things clearly: what needs to be done, which materials to rely on, what form the result should take, and what tone to use. When an answer is unsatisfactory, revise it item by item along these lines. OpenAI's prompting guidance for ChatGPT lists "be clear, be specific, provide enough context, and avoid ambiguity" as basic practices, and describes prompt engineering (designing and refining the input so the answer moves in the direction you want) as an iterative process rather than something you get right in one pass. A separate API prompt engineering document adds a premise that is often overlooked: model output is non-deterministic (the same input is not guaranteed to produce exactly the same result every time).

What four things should a prompt cover at minimum

Goal. Start with a verb and state who the finished result is for. "Take a look at this material for me" becomes "Compress this material into three points for reading aloud at a meeting," and the model knows what to do.

Material. Paste it directly into the conversation and explain how to handle what is missing, for example "mark anything not written down as 'to be confirmed' and do not fill it in yourself."

Result format. State the length, the structure, and the items that must appear. Being clear here saves you from repeatedly asking for "more revisions" afterwards.

Tone. Put it on its own line, using specific adjectives.

Suppose you want to turn a set of notes into a meeting notice. You could write:

Goal: Turn the notes below into a meeting notice for colleagues. Material: Use only the information in the notes; mark anything not written as "to be confirmed" and do not fill it in yourself. Result format: Within 200 characters, in four points, covering the time, place, topics, and what each person should prepare. Tone: Formal and concise. Notes: Wednesday 3 p.m., third-floor meeting room, discussing next quarter's schedule, Xiao Li to prepare the data.

The most direct benefit of writing them separately is that when the result is wrong, you can pinpoint which line is at fault instead of rewriting the whole thing.

Why the tone should be written as specific adjectives

Official guidance points toward using descriptive adjectives to specify tone, with examples including formal, informal, friendly, professional, humorous, serious. Phrases like "write it more professionally" or "make it more natural" do not point to a specific style, so the model can only guess. Replace them with one clear adjective and add a sentence with the example you want, and you have far more room to compare and judge.

One thing to note: this approach addresses whether the model can understand your requirements; it does not mean the content is necessarily accurate. For facts, numbers, and times, you still need to check them against the original source yourself.

When the answer is wrong, first work out which category of problem it is

Official guidance lists iterative improvement as a basic practice: give a first draft, look at the answer, then adjust the wording, add context, or simplify the request based on the output. In practice, change only one thing at a time and attribute the problem to one of three categories:

  1. The goal is too vague. The answer is broad and generic, or misses the point. Replace the request with a verb plus a purpose, for example "compress it into three paragraphs for reading aloud at a meeting."
  2. The material is not in the prompt at all. Just replying "be more accurate" will not change anything. Paste the original text or data in first, and note which fields must be preserved.
  3. Two requirements pull against each other. You want it finished in one sentence and also want every detail covered. Split it into two steps: list the key points first, and after you confirm them, compress.

Change three or four things at once and even if the result improves, you cannot tell which one made the difference, and with a different type of task you have to start experimenting from scratch.

Different models call for different emphasis in how you write

Official documentation distinguishes two types of models: reasoning models generate a thinking process internally, are good at complex tasks and multi-step planning, but are generally slower and more expensive; another type of model aimed at general text is faster and cheaper, but depends more on you stating the execution requirements explicitly. The documentation also notes that different model types, and even different versions in the same series, may perform differently on the same prompt.

In terms of how you write: for tasks that require multi-step judgment, spell out the steps and the decision conditions; for tasks like rewriting, compressing, or translating, getting the format and boundaries clear is usually more important. When someone says "this prompt works well," it only means it works at their entry point; it cannot be taken directly as a conclusion that it applies to yours as well.

There is a boundary to draw here: the instructions parameter, the developer and user message roles, fixed model versions, and test and evaluation suites in that document are concepts intended for developers building applications with the API (an application programming interface, a service that lets programs call the model). The ChatGPT chat window does not have these switches, so do not treat "setting a developer role" as something you do in chat.

A few recurring mistakes

  • Cramming five or six requirements into one paragraph, so when something goes wrong there is no way to start, and it is hard to judge which sentence to change.
  • Using "make it more professional" or "make it more detailed" instead of adding material, when the problem is actually the material, not the wording.
  • Asking for both brevity and complete coverage, two conditions that conflict in themselves.
  • Judging "this prompt doesn't work" on the basis of one unsatisfactory answer. Output itself involves randomness, so it is worth trying two or three times to see whether it consistently deviates from your requirements.
  • Treating the background in your own head as information the model already knows, then blaming the answer for being inaccurate when it was never written into the prompt.

Writing prompts does not require memorizing templates; what it requires is breaking the requirements apart and changing only one thing at a time. If you are using it for the first time, you can start with the free plan beginner's guide and practice on a small task that is easy to check.

Topic: Account and usage(All articles in this topic)

Ken|Operations Lead at TopupCat

Content is compiled by the TopupCat content team and published under the name of Ken, Operations Lead.View author profile

Related reading

Related products: ChatGPT Go · ChatGPT Plus · ChatGPT Pro · Top-up guide · FAQ · Self-service troubleshooting