How to write an effective prompt: the 5 rules that actually change the answers
How you phrase the request is the part that is up to you, and the part that changes the answer most. Five practical rules, with before-and-after examples.
- Level
- beginner
- Time
- 5 minutes
- What you need
- Nothing, just curiosity
A prompt is the message you write to an assistant based on artificial intelligence, nothing more. Plenty of people treat it like a Google search: three words and go. And the answer that comes back is generic, about as generic as the question. The five rules below work with any assistant, and once you apply them they make the difference between a mediocre reply and a genuinely useful one.
Prompt: the message you write to the assistant. It is the only thing the program knows about you and your problem, which is why it matters so much. Context: everything you have written earlier in the same conversation, which it keeps in mind while answering. It has a maximum capacity, though: in very long conversations the oldest parts drop out of its reach, and from then on it is as if it had never read them.
1. Give context before the request
The model does not know who you are or why you are asking. Before: “Write an email to the client.” After: “I’m an IT consultant. I need to email a client who paid 40 days late, keeping a friendly tone because I want to carry on working with them.” The second version gives you back an email much closer to the one you would have written yourself, and often a light touch-up is all it needs.
2. Specify the format you want
Do you want a list, a table, a 100-word text, a JSON file? Say so. Writing “answer with a three-column table, product, price, pros and cons” saves you a second request just to put everything back in order.
3. Show an example
If you have a style in mind, paste an example and ask for it to be imitated. “Here’s how I usually write product descriptions: [example]. Write the description of this new item the same way.”
One example says more than ten adjectives.
4. Ask it to reason step by step
For logic problems, calculations or decisions with several variables, asking it to show the steps and then check the result is nearly always useful. Not so much because the model “thinks better”, but because a written path is something you can verify yourself, and mistakes surface where before there was only a conclusion. One request that works well is this: “list the steps, then check the result and tell me whether it holds”. On newer models, which lay out their steps on their own anyway, the gain is smaller, though the final check stays handy.
5. Iterate, don’t start over
The first answer is a draft. Instead of rewriting everything, correct it: “Shorter”, “Drop the second paragraph”, “Use a less formal tone”. The model keeps the conversation in mind, and two or three corrections are usually enough to sort things out.
If after a few rounds the answer gets worse rather than better, open a new conversation and paste in the best version you got. Starting clean costs less than pushing on.
A common mistake to avoid
Avoid asking “can you do X?”. These days it almost always does it anyway, but every so often it answers yes and stops there. Write “do X” and there’s no doubt, and it costs you nothing. And if the answer contains numbers, dates or quotes, check them one by one: assistants can invent details with great confidence.