The pattern in both wins is the same: you supply every fact, the model supplies the shape. When you keep to that division, the output is trustworthy because there is nothing in it you did not put there. When the model starts adding facts of its own — and it will, if you leave room — you have to check every one.
Win 5: notes into a document
Try it
Five bullet points into a team update
- Did it work? Count the facts: 8%, 4%, 5%, Priya, 6 October, November, the conference budget. All seven should be there and nothing else should be. A memo that says the redesign was delayed "due to resourcing" has invented a reason.
- If not: the fix is the same as lesson 1 — *"Use only the facts in the notes."* You will find yourself typing this sentence a lot. That is fine; it works.
Win 6: compare options
Give the options, the criteria and what you will use the thing for. Ask for a table and a one-line recommendation. The table forces the model to say something about every option on every criterion, which is where vague comparisons usually hide.
Try it
Three laptops for one family
- Did it work? All three options appear in every row, the recommendation is one line, and there is a stated condition ("if the video editing became daily, choose C"). If a row says "varies", the model dodged — ask it to commit.
- Watch for: any specific claim about a product — a chip name, a battery hour figure, a price — that you did not give it. Those are the model's assumptions, and they need checking against the maker's page before you spend $1,500 on them.
| You supply | Model supplies | Danger zone |
|---|---|---|
| The options and their prices | The table structure | Prices you did not give it |
| What you will use it for | Which criteria matter most for that use | Specs it assumes about a product |
| Your constraints (budget, who uses it) | A recommendation that fits them | A recommendation that quietly ignores one |
Knowledge check
The laptop comparison says option B has "excellent 14-hour battery life". You never mentioned battery. What should you do?