How to give AI clear instructions that get a reliable result
The previous two lessons covered where AI helps and which of your data is safe to feed it. The third skill decides output quality more than anything else: how you ask. The same tool can hand you an excellent, ready-to-use result or a generic one that will not do — and the difference is usually not the tool, it is the clarity of what you asked. This lesson teaches you to build the instruction that cuts guesswork and gets you a result you can use as-is.
The rule: the tool does what you asked, not what you meant
AI does not know your business context, cannot see what is in your head, and will not assume what you did not say. It reads only the words you wrote and builds on them. When you give a vague request like "summarise this," you leave it to guess the length, the style, the purpose and the audience — and it will usually guess wrong. The rule is simple: the more you move from your head into the instruction explicitly, the less it guesses and the closer the result gets to what you actually want.
A weak output is usually not the tool failing, but an incomplete instruction. Fix the instruction before you blame the tool.
Four elements of any practical instruction
Almost any useful request is complete with four elements. Missing one is the most common cause of an output that will not do, so make them your mental checklist before you send:
| Element | The question it answers | Example |
|---|---|---|
| Context | Who are you and what is your situation? | I own a small retail shop in Riyadh |
| Task | What exactly do you want? | Classify these complaints by problem type |
| Material | What input does it work on? | The complaints text listed below |
| Output format | How do you want the result? | A two-column table: problem type and count |
Notice the four elements need no technical language or special phrasing — just plain sentences describing your situation and your request. Whoever writes those clearly gets ahead of whoever hunts for magic words.
A concrete example: from a vague request to a usable result
The owner of a family restaurant wants a message announcing changed working hours in Ramadan. The vague request:
Write a message about the change of hours in Ramadan.
That yields a generic text whose tone may not fit the restaurant or the channel. The request completed with the four elements says:
I own a family restaurant in Riyadh. Write a short customer message announcing that Ramadan hours run from after the Maghrib prayer until midnight. Use the times listed below without changing them. Make it a single message in a friendly tone, no more than three lines, ready to post directly on WhatsApp.
The second request usually gives a ready-to-use result on the first try, because it left nothing to guess: it set the context, fixed the task, pinned the material, and drew the shape, length and channel of the output.
Improve by iterating, not by padding
You will not always get it right the first time, and that is normal — it does not mean the idea failed. Better to fix the instruction step by step than rewrite it whole: if the output is too long, say make it shorter; if the tone is too formal, say soften it; if it missed a point, say add it. A small, targeted tweak is faster and more precise than a longer new request — and it is what an experienced user does.
Common mistakes that spoil the output
- One request carrying three mixed tasks; split them into sequential ones.
- Leaving length and format to the tool, then disliking the choice you never specified.
- Assuming the tool knows your internal jargon without explaining it.
- Accepting the first output without a critical read; a good instruction improves the result but does not replace your review.
Checklist before you send
- You set your context: who you are and the nature of your business.
- You defined one clear task, not several mixed together.
- You attached the material the request works on.
- You described the output shape, length and channel.
- You intend to read the result critically, not accept it as-is.