Checking
Eight rules, a nine-point checklist and prompts to have the AI check its answer.
To remember
The 8 rules
- Context first, then the question.Country, rules, industry, people involved: the difference between a right answer and an answer that is right for someone else.
- Ask for a draft, not the truth.What comes back is material to check, not a decision.
- Say what you do not want, too.Phrases to avoid, forbidden shortcuts, the wrong tone: limits matter as much as requirements.
- Do not take the constraint for granted.If a deadline, budget or premise could be wrong, ask the AI to check it.
- Demand the doubts.Ask it to mark [to be verified] and to list the assumptions it made.
- Be wary of precise numbers with vague sources."68% of managers" without an author, an organisation and a title is not data. Who says so? Where can I find it?
- Prepare the context once, in the project.Fixed instructions, materials without sensitive data, your method as a skill. Then the question can be short.
- You still sign off.AI speeds things up; responsibility, legal and ethical, cannot be delegated.
Before you use an answer
The nine-point checklist
Two minutes. One "no" is enough to stop and check.
After the answer
Have the AI check its answer
The checklist tells you what to check. These five requests get the AI to do it, right after an important answer. Copy them into the same chat.
Which statements in your answer are verifiable facts and which are assumptions? Mark them with [fact] and [assumption].
For every number you used, tell me where it comes from. If you do not have a source I can open, write [to be verified].
What is the strongest argument against your conclusion? Answer as someone who disagrees would.
What information are you missing to be sure of your answer? Ask me at most three questions.
Reread the answer as [a lawyer / a fiduciary / an operations manager] in Switzerland would: what would they challenge, and why?
An AI that checks itself helps, but it does not replace verification: you open the sources yourself.