A good AI prompt is specific and structured: it tells the model who to be, what to do, what context to use, what format to return, and what limits to respect. The single biggest improvement most people can make is to stop writing vague one-line requests and start giving clear instructions with an example of what a good answer looks like. Secret magic phrases matter far less than clarity. This guide gives you a reliable framework, a before-and-after example, and the mistakes that quietly ruin otherwise good prompts.
The five-part framework
Almost every strong prompt covers five things. You do not need all five every time, but naming them turns a vague ask into a precise one.
- Role — who the model should act as. "You are a careful copy editor." This sets tone and depth.
- Task — the exact thing to produce. "Rewrite this paragraph to be clearer and shorter."
- Context — the material and background it needs, pasted in or described.
- Format — the shape of the output. "Return a three-column table" or "answer in under 100 words."
- Constraints — what to avoid or require. "Keep the original meaning. Do not add new claims."
Before and after
| Weak prompt |
Strong prompt |
| Write about our product. |
Act as a marketing writer. Write a 60-word product blurb for busy small-business owners, focused on time saved. Plain language, no hype. |
| Summarize this. |
Summarize the text below in five bullet points, each under 15 words, capturing only the decisions made. |
| Help me with code. |
You are a senior Python developer. Explain why this function throws a KeyError and give a corrected version with a one-line comment. |
The strong versions are not longer for the sake of it. Each added word removes a guess the model would otherwise make. Specificity is what separates a usable answer from a generic one.
How to iterate
- Start with the framework. Cover role, task, context, format, and constraints in your first prompt.
- Read what went wrong, not just what is missing. If the tone is off, add a tone instruction. If it invented facts, tell it to use only the supplied text.
- Add an example. Showing one good sample output is often more effective than another paragraph of instructions.
- Break big asks into steps. Ask for an outline, approve it, then ask for the draft. Staged prompts beat one giant request.
- Save what works. Keep your best prompts as templates so you are not rebuilding them each time.
If you want grounded, checkable answers rather than confident guesses, pair good prompting with an understanding of how RAG grounds AI in real sources.
Common mistakes
- Asking vague questions and blaming the model for a vague answer.
- Copying "magic" phrases that sound powerful but change no actual instruction.
- Forgetting to specify format, then spending more time reshaping the output than writing the prompt would have taken.
- Dumping a huge unstructured wall of text with no clear task buried inside it.
- Expecting one perfect response instead of treating prompting as a short conversation.
FAQ
Do I need to learn prompt engineering?
Not formally. The fundamentals are clarity, context, and format. If you can brief a capable assistant well, you can prompt well.
Are there universal magic words that improve any prompt?
No. Phrases like "think step by step" can help on reasoning tasks, but they are no substitute for clear instructions. Specificity beats any phrase.
Should prompts be long or short?
As long as they need to be to remove ambiguity, and no longer. A short, precise prompt beats a long, rambling one.
Why do I get different answers to the same prompt?
AI output is partly random by design. For consistency, tighten constraints and format, and lower the creativity setting if your tool exposes one.
Where to go next
What RAG is and how it grounds answers, how to use ChatGPT for business, and how to use AI for writing.