A good AI prompt is not about finding a magic phrase. It is about giving the model enough useful context to understand the job, the result you want, and the limits it should work within. When an answer is vague, repetitive or aimed at the wrong audience, changing the wording of one sentence is often less useful than changing the structure of the request.
This guide shows a practical way to write prompts for everyday work. You can use the same approach across different AI assistants because the underlying problem is the same: reduce ambiguity before the model has to guess.
Start with the outcome, not the tool
Before writing a prompt, decide what you want to have at the end. “Help me with my report” leaves the task open-ended. “Turn these meeting notes into a one-page action list with owners and deadlines” gives the model a specific job.
A useful prompt normally answers four questions: What should be done? What information should be used? What should the result look like? and What should the model avoid?
A simple prompt structure that works
| Part | What to provide | Example |
|---|---|---|
| Task | The action you want | Summarize the report |
| Context | Who, what or why | It is for a school principal |
| Source | The material to work from | Use only the text below |
| Output | Format and level of detail | Five bullets and a short conclusion |
| Constraints | Important boundaries | Do not invent missing facts |
Weak prompt vs. useful prompt
Weak: “Write something about employee attendance.”
Better: “Write a short attendance policy summary for employees at a small office. Use plain English, cover late arrival, absence reporting and repeated violations, and keep it under 250 words. Do not invent legal requirements; mark any point that depends on local law.”
The second prompt is better because the model does not have to guess the audience, scope, length or legal certainty you expect.
Tell the AI what source material it may use
If accuracy matters, distinguish between facts supplied by you and general background knowledge. For example: “Use only the information in the document below. If the document does not answer a question, say that the information is missing.” This does not make the answer automatically correct, but it reduces the chance that a missing detail is silently filled with a plausible-sounding invention.
Specify the output before you get a long answer
Tell the model what a useful result looks like. A table may be better than paragraphs for a comparison. A numbered procedure may be better than an essay for a task. If you need text that can be pasted into another application, specify the format and any character or word limit.
Use examples when the style is difficult to describe
If you want a particular structure, give one small example rather than a long list of abstract instructions. For instance, provide one sample product description and ask the model to follow the same field order for the remaining products. Check that the example itself is accurate before using it as a pattern.
Use follow-up prompts instead of rewriting everything
You do not always need one enormous prompt. A better workflow can be iterative:
- Ask for a first draft.
- Point out what is missing or wrong.
- Ask for a revised version with those corrections.
- Ask the model to list assumptions it made.
- Verify important facts independently.
This is especially useful for complex work where you cannot describe every requirement in the first message.
Useful correction prompts
- “Keep the facts, but remove repetition and shorten the answer to 300 words.”
- “Separate facts stated in the source from assumptions.”
- “List the questions you need answered before you can complete this accurately.”
- “Check the calculation again and show the inputs you used.”
- “Rewrite this for a beginner without changing the meaning.”
Common prompting mistakes
- Giving a vague task and expecting the model to infer the goal.
- Adding conflicting instructions in different parts of the prompt.
- Providing too much irrelevant background.
- Assuming a confident answer is a verified answer.
- Asking the model to “never make mistakes” instead of giving it source material and a verification process.
- Putting private information into a prompt without considering whether the service should receive it.
A practical prompt template
You can adapt this structure:
Task: [what I need you to do]
Context: [who this is for and why]
Source: [information you should use]
Output: [format, length and level]
Constraints: [what to include, exclude or flag]
If information is missing: [what you should do]When a better prompt still will not fix the answer
Prompting cannot compensate for missing evidence. If you ask an AI system for a current legal rule, live price, medical decision or other high-stakes fact without giving it a reliable source or using an appropriate current-data workflow, a beautifully written prompt can still produce a wrong answer. Treat prompt quality as one part of the workflow, not as a guarantee of accuracy.
Quick checklist
- □ Is the task specific?
- □ Does the AI know who the result is for?
- □ Did you provide the relevant source material?
- □ Did you specify the desired format?
- □ Did you state important limits?
- □ Did you tell it what to do when information is missing?
- □ Will you verify facts that matter?
Quick answer
The best AI prompts make the task, context, source material, output and constraints explicit. Start with the result you need, remove ambiguity, work in smaller iterations when appropriate, and verify important claims rather than treating a polished response as proof that it is correct.
Related Tervilo resources
After improving a prompt, use the relevant Tervilo tools and guides to check calculations, documents or other outputs instead of relying on the AI response alone.