What it is
The model has not seen your customer’s contract, your Q3 numbers, the decision your team made on Tuesday or the reason the launch slipped. Giving it the facts means putting that material in the prompt: your notes, the figures, the email thread, the paragraph of the policy that applies. On the ten checks this is Material, the third check.
It is also the check most prompts miss. In the thinkhard.si study of 319 one-line work requests, almost none gave the model the facts it needed. The study names that gap, together with the missing instruction for what to do when a fact is absent, as the source of most of the wrong numbers and made-up details people complain about.
The facts do not need to be written up. Rough bullets are fine, and often better: the model writes the sentences, and you can see at a glance what it was given.
When it helps
It helps with anything specific to your organization or your situation, which is most work writing: status updates, follow-up emails, performance reviews, proposals, bios, replies to customers. The test is simple. If a new colleague would have to ask you for it before starting, paste it.
It helps most where a wrong detail is expensive. A made-up figure in a board update or a promise the company never made in a customer reply costs more to undo than it would have cost to paste the real one. When the facts are in the prompt, checking the answer is also faster, because you are comparing it with a list you wrote rather than with your memory.
When it does not
For general knowledge, the facts are already in the model. Explaining a well-known concept or writing a standard spreadsheet formula needs no pasted material. The exception is anything recent or changeable, such as prices, laws and product details, where the model’s knowledge may be out of date: paste the current source.
More is not better. A whole wiki page pasted to answer one question buries the paragraph that matters and costs more to send. Paste what the answer depends on; for long material, the Token Counter shows the size, and putting long documents first helps the model use it.
Facts alone do not stop invention. The model still fills the gaps between them, so a review with three achievements may come back with a fourth. Add “use only these facts” and say what to do when one is missing.
Some facts should not be pasted at all, or not into a personal account. Customer records, salaries and HR details belong in the tool your employer approved, or go through Safe Paste first so names and numbers become placeholders. And pasted text can carry instructions of its own, such as a line in an email that says “reply by Friday”, which is one reason to fence it off.
Before and after
Two requests where the after version adds only the facts. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write a follow-up email to a prospect after our demo.
Write a follow-up email to a prospect after our demo. Facts: - Demo on Tuesday with the operations director of a 40-person delivery company. - They liked route planning. They worried about the six-week setup. - We offered a 30-day pilot at no cost, with setup done by our team. - They suggested a call with their IT lead next week.
With nothing to go on, the model writes a follow-up that could go to any prospect after any demo, recapping features the prospect may never have seen. With the facts, it can answer the setup worry with the pilot offer and propose the call with their IT lead, which is the email you would have written yourself.
Write a performance review for my report, who had a good year.
Write a performance review for my report, who had a good year. My notes: - Led the move to the new ticketing system in March, two weeks early. - Cut average first-reply time from 9 hours to 3. - Trained the two new hires in June. - Still hands off escalations late; we agreed to work on this in the spring. - Rating I am giving: exceeds expectations.
The first version produces praise that fits anyone who had a good year, and it may invent an achievement to fill the page. The second can cite the March migration and the drop in reply time, and it has the development point to include, so the review sounds like it came from someone who watched the work.
The scores move by one or two points because Material is one check of ten; the first pair also passes the estimate’s Scope check by accident, since “30-day” looks like a length. The answers move more than the scores. The guide to performance reviews with AI covers the rest of that prompt.
The line to add
Facts:
- {{fact}}
- {{fact}}
- {{fact}}
Use only these facts. If you need one that is not here, ask me.
The last line belongs with the list. Without it, the model treats your facts as a starting point; with it, they are the boundary. List a fact even when it seems obvious to you, such as the date, the product name or who agreed to what. Obvious to you is exactly what the model cannot see.
In ChatGPT, Claude, Gemini and Copilot
All four vendors ask for this. Microsoft’s guide to Copilot prompts calls it the source, the information or samples Copilot should use, and says it could be a file you attach, a web page, or your work emails, chats and files. In Copilot at work, pointing to the right file can replace pasting it.
Google’s guide for Gemini has a section on adding context: include the information the model needs instead of assuming it already has it. Google’s Workspace prompting guide says to provide as much context as possible. OpenAI’s prompt engineering guide, for the models behind ChatGPT, describes adding context to give the model access to proprietary data it was not trained on.
Anthropic adds the second half. Its guide to reducing hallucinations suggests telling Claude explicitly to use only the provided documents and not its general knowledge, which is the “use only these facts” line above.
Key takeaways
- The model knows the world, not your week. Paste the figures, names, dates and decisions.
- If a new colleague would have to ask for it, it belongs in the prompt.
- Rough bullets are enough. The model writes the sentences.
- Paste what the answer depends on, not everything you have.
- Pair the facts with “use only these” and an instruction for what is missing.
Questions
Why does ChatGPT make things up?
Often because it was asked for something it did not have the facts for. A model writes the most plausible text it can, and when the prompt leaves out the real numbers or names, plausible ones fill the gap. Paste the facts, tell it to use only those, and tell it to ask when something is missing. The guide to reducing hallucinations goes further.
How much context should I give AI?
Everything a new colleague would need to do the task without asking you, and not much more. That usually means the specific facts, such as figures, names, dates and decisions, plus the source text the answer depends on. A paragraph that matters does more than a folder that might.
Is it safe to paste work documents into ChatGPT or Claude?
It depends on the document and the account. For anything confidential, use the workplace version your employer has approved. When you are not sure, swap names and numbers for placeholders first with Safe Paste, and the guide to anonymizing text shows what to remove.
Prompts to use with this
- Explain why two reports give different numbersTwo sources disagree on a figure, such as a market size, a rate or a survey result, and you need to know why before you choose one.