11 Lesson 11 of 15
Asking as specification, not instruction
The biggest improvement most people can make: stop briefing it like a colleague, and start filling it in like a form.
If I explain clearly, it will do what I want. You wrote a clear, polite, complete request. A good assistant would have understood it. So if the result is wrong, either the tool is not very good or you were not clear enough, and the fix is to explain again, more nicely.
What helps is to think of a prompt as a specification you keep improving, not an instruction you give once.
See it in the lessonThis is the most closely studied problem on the whole site. Researchers gave non-experts a tool for writing prompts and watched what they did. Two things kept going wrong. First, people treated it like instructing a person: they assumed it shared their background and would work out the point. Second, they read too much into single results, treating one good or bad answer as proof about the whole approach. They tried things at random: tweak, look, tweak, look, with no plan.
What helps is to think of a prompt as a specification you keep improving, not an instruction you give once. Three moves do most of the work.
Three moves
1. Say what the output should look like
Instruction
Can you help me summarise this report?
Specification
Summarise this in five bullet points, each under fifteen words, aimed at someone who has not read it. No preamble.
2. Give one example of what you want
One example is worth several paragraphs of description, because it pins down things you would not think to say: tone, length, how formal, how much detail. If you have a good version from last time, paste it in and say “in this style”.
3. Say what you actually want it for
Instruction
Write an email declining the invitation.
Specification
Write a short email declining a conference invitation. I want to be asked again next year, so warm but clearly final. Three sentences. No apology for the delay.
Then revise, and revise the prompt
When the result is wrong, the urge is to reply “no, not like that” and explain. That sometimes works, but it buries the fix in a chat you cannot reuse. A better habit is to go back and change the prompt itself, add what was missing, and run it fresh. Then you have something you can use again. That is the difference between getting one good result and getting good results every time.
And remember the lesson on repeat answers. If you change your prompt and the result improves, that might be your change, or it might just be a different draw. If it matters, run the new prompt twice.
Why this works is easier to see than to say. Picture every reply that would be a reasonable answer to Can you help me summarise this report? There are a great many. Each thing you specify rules most of them out. Switch the specifications on and watch the cloud shrink towards the reply you actually wanted.
Read this diagram as text
There are 640 possible replies, varying in format, length, audience, tone and detail. Stating the format keeps a quarter of them. Stating the length keeps a quarter of those. Saying who it is for halves what is left. Giving an example fixes the tone and most of the detail. With all four on, two replies out of 640 remain. The line under the buttons gives the exact count after each switch, and Run it describes one reply picked from those still allowed.
Build a specification
Fill in as many of the boxes as you like. The prompt assembles itself as you type, and you can copy it straight into whichever tool you use. Nothing you type leaves this page.
One analogy. It is closer to filling in a form than asking a favour. A form makes you state the things that are easy to leave out.
Where the analogy breaks: a form rejects you when a field is missing. This will happily make up a value for every field you left blank, and not mention it.
Take a prompt that gave you a disappointing result. Do not explain yourself to it. Add one example of what you wanted to the original prompt, and run it again from scratch.