AI principles · Chapter 5 of 8
Saying what you want
Where a request leaves a gap, the model often fills it with a likely guess, so a good request describes the result, the way to work and the behaviour you want, with the facts and an example only you can give.
Goals
- Read a vague request and the answer it got, and mark what the answer had to guess.To do
- Add parts to the request one at a time until the checker finds no gaps.To do
- Compare the answers to the vague and the full request.To do
- Pass the check.To do
Step 1 of 4: The read
A gap gets a guess
Chapter 4 was about what to hand over. This one is about how to ask. What you type to an AI chat assistant is a prompt: the text the model starts from, as in chapter 2. In a plain chat, with no access to your files or calendar, all the model knows about your situation is what you write.
Where your request leaves a gap, the model often fills it with what is likely in general, the way it picks likely words. The result reads well, so the guesses are easy to miss: a date nobody set, a promise nobody made.
Three things to describe
Write your request as a brief: the instructions you would give a capable helper who knows nothing about your situation. Three things are worth describing.
The result: who it is for, its form (a list, a notice, an email) and how long it should be.
The way to work: the steps to follow, and what to ask you before starting. “If a fact is missing, ask me” turns a guess into a question.
The behaviour: how it should sound, and what to avoid, such as promises you cannot keep.
What only you know
Two more parts carry what the model cannot guess. Context is the background facts: dates, names, what has already been decided. An example shows what good looks like, such as a notice that worked last time. Say what to take from it, such as its layout, or the model may copy its old date too.
Words like “make it good” add length, not information: they do not say what good means.
In the visual
You start from a vague request for a community garden notice and the answer it got, and mark what that answer had to guess. Then you add parts one at a time while a checker flags what is still missing. The checker is a few fixed rules that run in this page and look at which parts your request has; it is not an AI. Last, you send the full request and compare the two answers. The garden and both answers were written for this chapter to show the kind of guesses a model can make.
Step 2 of 4: Try it: Brief builder
Brief builder
Mark what a vague request left to guess, add parts until the checker finds no gaps, then compare the two answers.
Mode:RealRe-createdReal: the checker that flags gaps runs in this page, as a few fixed rules that look at which parts your request has; it is not an AI. Re-created: the garden, the request parts and both answers, including the question the assistant asks, were written for this chapter to show the kind of guesses a model can make. No model runs here, and nothing you do leaves the page.
Step 3 of 4: Check
Check
Next: Question 2 of 4
Step 4 of 4: What you can do now
What you can do now
- Explain why a vague request gets an answer full of guesses.
- Describe the result, the way to work and the behaviour you want from an AI.
- Add the context and an example only you can give, and say what to take from the example.