How to Write Better AI Prompts: A 5-Part Framework
- A good prompt states the goal, the audience, the limits and the output format.
- Specific details remove guesses; length alone does not.
- Treat the first answer as a draft and refine it with one change at a time.
- Save prompts that work, with a note about what they are good for.
Most disappointing AI answers are not a sign of a weak tool. They are the predictable result of a vague request. A language model has no access to your plans, your audience or last week's attempt. It only has the words you type, so every detail you leave out becomes a guess, and the model fills each guess with the most average option available. That is why a short, general prompt produces a short, general answer.
This guide gives you a five-part structure you can apply to almost any AI assistant, plus a template, a troubleshooting table and a worked example of improving one prompt over three rounds.
The five parts of a strong prompt
1. Goal: what you want and what it is for
Start with one sentence that includes the task and its purpose. "Write a caption" is a task. "Write a caption for a 20-second video that shows a quick desk-setup tour for beginner creators" is a goal, because it tells the model the format, the topic and who will read it. If you cannot finish the sentence "this is for...", pause and decide that first.
2. Context: facts the model cannot guess
Give the two or three facts that change the answer: the audience, the platform and the tone. Skip background that does not affect the result. A paragraph about your channel's history rarely helps, while "my viewers are beginners who edit on a phone" helps a lot.
3. Constraints: limits that make the answer usable
Constraints are the difference between an answer you can paste and an answer you must rewrite. Set a length ("under 100 words"), a structure ("three bullets, then one question"), a style ("plain words, no jargon") and a list of things to avoid ("no hashtags, no claims about income").
4. Example: show the style you want
One sample is often worth a paragraph of description. Paste a short piece of text in the voice you like and write "match this tone, but do not copy the wording." Examples work especially well for voice, formatting and naming patterns.
5. Output format: how the answer should arrive
Say whether you want a numbered list, a table, a script with time stamps or plain paragraphs. If you plan to paste the result into a spreadsheet or a video editor, ask for exactly that layout so you spend less time cleaning it up.
Weak versus strong: three side-by-side examples
| Task | Weak prompt | Strong prompt |
|---|---|---|
| Video ideas | Give me video ideas. | I teach beginners to edit on a phone. Give me ten video ideas, each with a working title and a one-sentence hook. Keep each idea filmable in under 30 seconds. Return a numbered list. |
| Caption | Write a caption for my video. | Write three captions for a video that shows how to stabilize shaky phone footage. Audience: beginner editors. Under 30 words each, friendly tone, end with a question. |
| Summary | Summarize this article. | Summarize the article below for a busy creator in five bullet points. Keep each bullet under 20 words and mark any claim that needs a source check. Article: [paste text]. |
The strong versions take under a minute longer to write. In return, they remove most of the follow-up messages you would otherwise need to correct length, tone and format.
A reusable template
Goal: [what you want, and what it is for]. Audience: [who will read or watch it]. Platform: [where it will be used]. Tone: [two or three words]. Constraints: [length, structure, things to avoid]. Example of the style: [optional short sample]. Output format: [list, table, script or paragraphs].
A worked example: improving one prompt in three rounds
Round 1, the first attempt. "Write a script about phone editing." The answer is a generic, rambling script with no clear audience. The problem is not the model; the request contains no audience, length or purpose.
Round 2, add goal and constraints. "Write a 30-second script for a short video that teaches beginners one phone editing trick. Start with a one-line hook, then three short steps, then a closing question." Now the structure is right, but the trick is random.
Round 3, add context and an example. "The trick is trimming dead air at the start of a clip. My tone is calm and practical, like this sample: 'Most beginners leave two seconds of silence at the start. Here is how to cut it in ten seconds.'" The result now matches the topic and voice, and only light editing is left.
Notice that each round changed one thing. Changing five things at once makes it impossible to know which change helped.
Troubleshooting table
| Symptom | Likely cause | Fix |
|---|---|---|
| Answer is generic | No audience or example | Add who it is for and paste a short sample |
| Answer is too long | No length or structure limit | Set a word limit and a number of bullets |
| Tone feels wrong | Tone never specified | Name the tone or paste a sample sentence |
| Part of the request is ignored | Too many tasks in one prompt | Number the requirements or split into two prompts |
| Facts look doubtful | Model is guessing | Ask it to flag uncertain claims, then verify them yourself |
| Results vary a lot between runs | Prompt leaves choices open | Fix format and constraints; ask for a fixed structure |
Common mistakes and how to fix them
- Writing a wall of text. Long prompts are fine when every sentence changes the answer. Cut anything that does not.
- Asking for several different outputs at once. Ask for one deliverable per prompt, then combine them yourself.
- Never saying what to avoid. If you do not want emojis or hype words, say so.
- Accepting the first answer. Reply with specific feedback such as "shorter, and lead with the benefit."
- Copying output without checking. Names, numbers and claims must be verified before publishing.
When this does not work
A better prompt cannot give a model information it does not have. If you ask about a recent event, a niche tool or a private document, the model may invent details with great confidence. In those cases, paste the source text into the prompt and ask the model to use only that text, or look the answer up yourself. Structure also cannot fix a weak idea: if the goal is unclear, a perfectly formatted prompt will still return a mediocre answer.
Build a personal prompt library
When a prompt works, save it with three notes: what it is for, which tool you used, and what you changed to make it work. A simple document or spreadsheet is enough. Over a few weeks, you will have a set of proven starting points, and writing new prompts becomes a matter of editing an old one.
Frequently asked questions
How long should a prompt be?
As long as it needs to be, and no longer. For simple tasks, two or three sentences with a clear goal and format are enough. Long prompts help when each part changes the result, for example when you include a sample of your style.
Do the same prompts work in every AI tool?
The structure works in most assistants, but results differ between tools and versions. Test your prompt in the tool you actually use and adjust the wording for that tool.
Should I tell the AI to act as an expert?
Role statements can nudge tone, but they do not add knowledge. Specific context, constraints and examples usually help more than a job title.
How do I stop the AI from making things up?
You cannot remove the risk completely. Paste your source material, ask the model to use only that text, ask it to say when it is unsure, and always verify facts before publishing.