
You ask for “a plan” and receive a long table that does not fit your resources or dates. The response is organised, but the information that would make it useful was absent from the request.
A good prompt explains the intended result and supplies the material needed to create it. Complex wording matters less than clarity about the task and what the assistant must not assume.
A prompt is the request or instruction you give an AI tool. Improve it by stating the result, relevant information, output format and constraints, then reviewing and refining the response with specific feedback.
On this page
What is a prompt?
It can be a short question, an image description, a revision request or a task involving a file. Length alone does not indicate quality. “Halve this paragraph without removing the conditions” may be enough when the paragraph is available.
The official OpenAI prompting guide explains starting in natural language and refining through conversation. Details that change the answer are more useful than a long introduction about the assistant’s role.
Four useful parts of a request
Name the action: explain, compare, organise, design or revise. Add essential information such as the audience, source and constraints. Specify the output if you need a table, short paragraph or execution list.
Use only the parts that help. For a larger task, define stages or ask for essential missing information before execution.

A writing template using approved facts
Write [content type] about [subject] for [audience]. Use only these facts: [facts]. Deliver [length or format]. Follow [a short style description or example]. Preserve [fixed details]. Ask about missing information instead of adding it.
For a service description, provide the actual scope and deliverables. Do not let polished marketing language introduce a guarantee you cannot support.
A sourced-research template
Research [specific question] for [market or country] during [relevant period]. Start with official sources. Separate published information from interpretation and link each important factual claim to its source. Identify gaps where evidence is insufficient.
Open the references and check dates, geography and measures. Use the available search tool when freshness matters; writing “research” does not establish that a search occurred. See the official search guide for ChatGPT.
An image brief for a design
Create a scene featuring [main subject] for [design use]. Use [aspect ratio]. Place the subject at [position] and leave space for copy at [position]. Lighting: [description]. Colours: [palette]. Avoid [unwanted details].
For example, place a notebook and pen on a white surface with the objects on the right and quiet copy space on the left. Define the destination before collecting visual details. The AI image guide covers review and finishing.
A table or notes template
Turn this material into a table with [columns]. Preserve the supplied values and do not infer missing information. Write “not specified” where needed. After the table, identify rows requiring review: [material].
For calculations, state the formula, unit and period. Check a sample yourself and request row or page references when they help you compare the output with the source.
Refine an unsuitable response
Replace “I don’t like it” with a specific issue. “The opening is generic. Begin with the customer’s question, remove the first two paragraphs and retain the example” creates a change you can inspect.
When the revision moves closer to the goal, work on the remaining gap rather than changing the entire request. Keep an earlier version if a new revision loses useful detail.

Before and after: supply the missing information
A broad request is “Write a content plan”. A more useful request is “Suggest six topics for a training centre offering beginner Excel courses to office workers. Each should solve a practical problem. Include the idea, example and content format, without inventing a budget or expected results.”
The second request specifies the audience, service, number of ideas and output. A publishing plan would additionally need resources, dates and channels; topic suggestions alone do not complete it.
Avoid conflicting and unsupported instructions
Several conflicting requirements can obscure priorities. State what matters most, such as accuracy before brevity or preserving service conditions before making the tone lighter.
Do not request invented data simply to fill a table. If you need an example, label it as illustrative. Explain the role of supplied material rather than adding many files without a defined question.
Keep a small prompt library
Save a useful request with an example input and an acceptable result. Include its review date and the feedback that improved it. Replace task data next time rather than carrying over another client’s information.
Review templates as your work and tools change. A template should remain a starting point rather than a rule that obstructs a new task.
Frequently asked questions
Must I say “act as an expert”?
It can clarify a perspective, but it does not replace reliable information or acceptance criteria. State the actual task first.
Are longer prompts better?
Use enough detail for the task. Remove instructions that do not affect the result and add details whose absence would lead to a wrong assumption.
Can I reuse a prompt across tools?
Often, but account for available capabilities and output types. Research and image tasks can require different tool choices even within the same conversation.
Improve a request you already use
Identify missing context in an old request and add a way to judge the output. Compare the results. The ChatGPT beginner’s guide provides a first-conversation walkthrough.