Sheet 4
Checking a draft the smart way
AI Tools Field Guide
Not affiliated with OpenAI, Google, xAI or Anthropic; your organization’s rules come first.
Nobody reads everything twice. Check the right facts, fast, with the model doing the finding.

Why it invents
- It continues text, it does not look things up. A question about an instruction is usually followed by an instruction number, so it produces one, real or not.
- Its knowledge stops at a date, and web search is a periodic index, not the live internet.
- Numbers are predicted, not computed. Days between dates, totals, per-diem: it predicts the digits that usually follow.
Make the model do the finding
- References beside every claim. Ask for the page or paragraph after each item; an item with none is the one to check.
- A fresh chat as reviewer. Give it the draft and the source: “list every claim the source does not support.”
- Ask what it assumed. One line at the end of the prompt turns its guesses into a short list.
- Match the depth to the risk. A rewording: one read. A brief: the source open. A number you will act on: recompute it.
The tells
- A reference, date or name you did not supply. “Where in the attachment is that?” If it cannot point to it, it invented it.
- A citation you cannot open. Open it. A citation is a pointer, not proof.
- A number with no working. Ask for the inputs and recompute.
- An answer that is suspiciously complete. “What is missing from the source?”
Take away
Source in, references beside each claim, a fresh chat to find the unsupported ones, your own eyes on the numbers you will act on.