The rules my AI writes for my other AI
My AI writes rules for my other AI. Read together, they are a list of AI’s weaknesses.
I use Claude and ChatGPT together. When Claude asks ChatGPT for a second opinion, it writes rules into the request without being asked. I collected them from four of my projects. Each one guards against a way I have seen these models fail.
“Only what the brief states; do not invent.” Models fill gaps with confident fiction.
“Mark anything you cannot verify.” Models state guesses in the same confident voice as facts.
“Agreement is not valuable output.” Models tend to agree with whatever is in front of them.
“Every finding must cite” a specific record or rule. Vague findings cannot be checked.
“Concise, no preamble.” Long answers bury the claims I need to check.
“Do not rewrite the post. Name the sentence and the problem.” A rewrite brings in new wording nobody has checked.
And one rule that is not words at all: I keep the two reviewers apart, so agreement is at least not an echo.
I did not write any of these. I read every one afterwards and would have approved them all.
A line in a prompt is not a control. A model can ignore it or satisfy it on the surface. But as a list of risks, it is a good one, and each risk needs a control with an owner and a test.
Which of these risks do your controls cover, and how do you know they work?
- The author’s own AI working sessions, October 2026. Each quoted rule is verbatim from a request Claude wrote to ChatGPT, saved in the author’s records. One quote is shortened, as marked.
Scope note. Claude and ChatGPT are named because they are the tools the author uses. This is not a product review, and either model could play either role.
Written for compliance and risk readers in regulated financial firms. Informational, not legal advice; Mike Bidun is not a lawyer.