Give it what it needs to know
Before you ask, give the AI the facts that are obvious to you and unknown to it.
Prevents mistakes like "A reply written for nobody".
Here is a request I'm about to send. Don't carry it out. Tell me what's missing. [paste the request, with any material you planned to attach] Read it as a capable new colleague on your first morning. You know nothing about me or my work beyond what is written above. List up to five questions you would need answered before you could do this well. Put first the ones that would stop you. After each, say in a few words how the answer changes the result. Each must be a question I can answer in one sentence. Then, for the questions that would stop you, tell me what you would assume if I sent the request as it is. Don't ask for anything the request already says. Don't answer the request or write any part of it. No praise, no summary, no closing offer.
Before you ask, give the AI the facts that are obvious to you and unknown to it.
Imagine you run a small catering business. A customer asked for a quote last week and hasn't replied. You ask AI:
Write a follow-up message to a customer about their catering quote.
It gives you:
Dear Customer, thank you for considering us for your special occasion. We wanted to follow up on the quote we shared and would be delighted to be part of your celebration. Please feel free to reach out with any questions.
It's polite. It could have been sent by any caterer to anyone.
Here is what you knew and never typed:
- The customer is Mr Kulkarni. The function is an engagement on 14 December, for 150 guests.
- You sent the quote last Tuesday. He asked for Jain food at one table.
- You never hold a date without a 25% advance, and December is filling up.
- He found you through his sister, whose wedding you catered in March.
None of that felt like information. It's the background you live in. The facts you're least likely to type are the ones you find most obvious.
Try this on the last reply that let you down. What would a capable new colleague have written from exactly those words, on their first morning, having met nobody? If they'd have written the same thing, the request was the problem. That's good news, because you can fix a request.
Think of a desk, not a filing cabinet.
Everything the AI can use is on the desk at once: your request, anything you pasted, what was said earlier in this chat, and your standing lines from Lesson 2. It reads the desk fresh each time it answers. Engineers call this the context window.
There is no cabinet behind it. It can't go and fetch the Kulkarni quote, last month's message or "the format you used before". If it isn't on the desk, it doesn't exist.
The desk is also cleared between chats. A new chat starts with only what you put in it. Memory and project features in your tool do one job: they put something back on the desk at the start, so you don't have to. The names and limits of those features change often. What they do underneath doesn't.
5 to 10 minutes. You'll end with about five lines of background you can paste above a request you send often. The AI asks the questions. You answer the ones you can.
You need the request you saved in Lesson 3.
Here's a filled example, from the printing shop:
| The AI asked | The answer |
|---|---|
| Who are the customers? | Families ordering wedding cards. Most order once, on WhatsApp. |
| What do they worry about? | The date, more than the price. |
| What can't you promise? | A delivery date before the printer confirms it. |
Look at what it said it would have assumed. Those are the guesses it was making every time you sent the request without your lines.
"More details" has room for a second note about one piece of work. One note is enough for now.
Your background notes
About five lines on the people you write for. Paste them above a request so you stop explaining them each time.
One line each. Who they are, what they already know, what they worry about, what nobody may promise them. Put a date at the top.
More details (optional)
Names, where things stand, decisions already made. Put a date at the top.
Can you say what the AI can see when it answers, and what it can't?
Look at the last reply that was well informed and still not what you asked. Was there too little on the desk, or too much?
What is the signal that tells you to start a new chat?
Your background note stays saved in this browser. Lesson 6 and Lesson 12 use it again.
You can now ask clearly and give the AI what it needs. Lesson 5 is about the replies that sound right and aren't, and how to catch them before they reach anyone.
For now, remember:
If the answer depends on it, put it on the desk.