A prompt is just the instruction you give an AI tool. Most disappointing results come from prompts that leave the AI guessing. The major AI providers publish prompting guides, and they agree on a handful of habits.
1. Give it a role and an audience
“You are a bookkeeper explaining to a sole trader” gets a very different answer from a bare question.
2. Give it the material
Paste the policy, the email, the data. Ask it to use only that material and to say when something isn’t covered. This is the single biggest fix for made-up answers.
3. Be specific about the output
Length, format (table, bullets, email), tone, British English. If you want a table with certain columns, name the columns.
4. Show an example
One example of a good answer is worth a paragraph of description.
5. Ask it to check itself
“List anything you were unsure about” or “ask me questions before you start” surfaces gaps before they become mistakes.
A template to copy
You are [ROLE] helping [AUDIENCE].
Task: [WHAT YOU WANT DONE].
Use only the material below. If something isn't covered, say so.
"""
[PASTE MATERIAL]
"""
Output: [FORMAT, LENGTH, TONE]. British English.
Finally, list anything you were unsure about.Every use case in our library comes with a ready-to-use prompt built on this pattern.
Sources
- OpenAI, Prompt engineering guide Vendor documentation
- Anthropic, Prompt engineering overview Vendor documentation