Checks three and four: the premise and the counter-case

Check three: the premise. Read your own prompt again. Did it contain an assumption? Did you ask "confirm that" or "isn't it the case"? If so, the answer may be agreement rather than analysis. Re-ask as an open question in a fresh chat and see if the answer changes.

Check four: the counter-case. Ask the tool, in a fresh chat, for the strongest argument against the output's main conclusion. Not "are there any problems," which invites a polite no, but "argue against this as a sceptical expert would." Read what comes back for anything you had not considered. Sometimes it is nothing. Sometimes it is the thing that saves you.

These two checks catch reasoning errors, which are subtler than factual ones and often more expensive.

In practice. A programme lead had AI draft a recommendation to consolidate two suppliers. Check three found her prompt had said "given that consolidation will save money." Re-asked openly, the tool was less sure. Check four produced a counter-case about single-supplier risk that the original draft had not mentioned. The recommendation changed.

Here is a piece of AI-generated analysis I am considering relying on: [ paste ]. Argue against its main conclusion as strongly as a sceptical expert in this field would. Find the weakest step in the reasoning, the assumption it depends on most, and what evidence would overturn it. Do not be balanced. Make the case against.

At your desk. Take the same output. Reread your prompt for premises. If you find one, re-ask openly in a fresh chat. Then run the counter-case prompt. Write down whether anything changed.

Write it down.

• The premise I had built in:
• The strongest counter-case:
• What changed: