AI prompting
Get a useful answer on the first try, and know what to do when you do not.
Most bad answers come from a prompt that left something out, not from the model being incapable. This guide is the short list of things worth putting in, in the order they matter.
1. Ask for the outcome, not the method
People often write prompts that describe how to think: "act as an expert marketer", "use first principles", "think step by step". Role and method lines are weak levers. What actually changes the answer is being specific about the result you want and who it is for.
Weak: You are an expert copywriter. Write a good product description.
Better: Write a product description for a stainless steel water bottle, 60 words, for a shop page. The buyer is comparing three bottles and wants to know why this one is more expensive.
The second version names the length, the placement, the reader and the question the reader is actually asking. None of that required a persona.
2. Give it the material
A model cannot see your document, your inbox, or last week's meeting. Anything you do not paste in, it will fill in from a general sense of how such things usually go, which is exactly the part that reads as generic.
Paste the real material: the draft, the transcript, the error message in full, the three competing options with their prices. Length is rarely the problem. A prompt with a thousand words of real context beats a clever one-liner almost every time.
3. Show one example
If you know what a good answer looks like, show one. One worked example does more than three paragraphs of instruction, because it carries tone, length, structure and level of detail at the same time.
This works for anything with a repeatable shape: subject lines, commit messages, customer replies, meeting notes, product names. Paste one you already like and say "match this".
Two examples are better than one if they differ in a way you want it to notice. Beyond about three, you are usually better off describing the rule.
4. Name the shape of the answer
Say what you want back, in the form you want it. A table, five bullets, a paragraph with no headings, valid JSON with these four keys, a numbered list where every item starts with a verb. Models follow format instructions closely, and it saves you reformatting by hand.
The same applies to what you do not want. "No preamble, start with the first item" removes the opening sentence that restates your question back at you.
If you want to be able to use the answer in a spreadsheet or a script, say so at this point, not after you get prose back.
5. Make it check itself
For anything where accuracy matters, add a checking step to the same prompt:
- Mark the uncertainty. "Flag any claim you are not confident about rather than smoothing it over."
- Separate source from inference. "For each point, say whether it comes from the text I pasted or from your own knowledge."
- Ask for the gaps. "List what you would need from me to answer this better."
That last one is the most underused prompt in general use. It turns a confident wrong answer into a short list of the things you forgot to include.
6. When the answer is wrong
Do not argue with it in follow-up messages. A long correction thread drags the whole conversation along with it, and the model keeps anchoring on its first attempt.
Edit the original prompt and run it again. Add the thing that was missing, cut the instruction it over-applied, paste the material you left out. Treat the first answer as a report on what your prompt actually said.
Two specific failure patterns and their fixes:
- It is too general. You did not give it the material. Go back to step two.
- It is confidently wrong about a fact. It did not have the fact. Paste the source, or ask it to tell you what it is unsure of before you use anything.
7. The short version
A prompt worth reusing usually has four parts. Keep this as a template:
That is the whole method. Everything else is refinement.