Asking for what you actually want
In one line: most disappointing answers are not the model being stupid, they are a request that left out something the model had no way to know.
The four things
A request that works tends to have four parts. You do not need all four every time, and naming them is useful because when an answer comes back wrong you can usually point at the missing one.
Who it should be. "You are a careful technical editor" changes the whole register of the reply. Not magic — it narrows the enormous space of possible continuations down to the useful corner of it.
What it needs to know. The background it could not possibly have: who this is for, what has already been tried, what the constraints are. This is the one people leave out, and it is the one that costs the most.
What you want done. Be specific about the form. "Summarise this" and "pull out the three decisions and who owns each" produce very different things.
How it should come out. Length, tone, format. "Six bullets, no preamble" saves you an edit.
The single best habit
Give it more context than feels necessary.
A model guessing from three words is guessing badly. Paste the actual email, the actual error, the actual paragraph. People routinely describe a problem in one line when they could have pasted the thing itself — and then wonder why the answer is generic.
This is the whole reason AIOS sends your note along with your question. It is doing the highest-value habit for you, every time.
Things worth knowing
Ask for the thinking when it matters. "Work through it step by step before answering" genuinely improves hard problems. On easy ones it just makes the answer longer.
Say what you do not want. "No preamble", "do not apologise", "no bullet points" all work, and save you from editing the same thing out repeatedly.
Give it an example. One example of the output you want is worth a paragraph describing it.
Iterate rather than restarting. "Shorter, and drop the second point" is usually better than rewriting the question. You are steering, not re-launching.
Ask it what it needs. "What would you need to know to answer this well?" is a genuinely good move when you are not sure what you left out.
What to do with a bad answer
Do not just try again. Work out which part was missing:
- Generic and obvious? → it did not have enough context.
- Right shape, wrong content? → it lacked the facts. Give them.
- Wrong shape? → you did not say what form you wanted.
- Confidently wrong? → it did not know, and guessed. Give it the source, or use a tool that can look things up.
Try this
Take a request you would normally type as one line. Write it again with all four parts — who, what it needs to know, what to do, how it should come out. Compare the two answers. The difference is usually not small.
If you want to go further
---
Next: What to be careful about · All lessons