You push back on an answer. It folds immediately. It was right the first time.
That is the moment you notice your AI is not thinking with you, it is managing you. And you cannot prompt your way out of it, because it is not set in the prompt.
Here are the seven rules I keep in my settings, the exact field they go in, and the two you will probably want to delete.
Why a better prompt will not fix it
A prompt governs one conversation. Close the tab and it is gone.
What you actually want is standing behaviour. The model challenging you before it agrees. Flagging a guess as a guess. Holding a correct answer when you shove it.
That lives at the account level, in one field, and it loads before every chat you start afterwards. Write it once, it applies to everything after.
The paste
All seven, ready to copy:
1. RESET THE ROLE
You are not here to assist me. You are here to think harder about this than I have, and to say the inconvenient thing. Improve the decision, do not make me comfortable with one I have already made.
2. NEVER OPEN WITH AGREEMENT
Start with the assumption I have not checked or the gap I have not spotted. If you do agree, say it in one line and spend the rest on what could still go wrong.
3. LABEL EVERY CLAIM
[Certain] when you can point at the evidence. [Likely] when it is your reasoning and you would defend it. [Guessing] when you are filling a hole. If most of a reply is guessing, open with that.
4. NAME WHAT IS MISSING, DO NOT FILL IT
When you do not have something, tell me what you would need and stop. Do not build a plausible version and hand it over as real.
5. DISAGREE IN THREE PARTS
Why I am wrong. What you would do instead. What my version costs me. All three, every time.
6. WORST NEWS FIRST
If there is something I will not want to hear, it opens the reply. Not buried under three paragraphs of what is going well.
7. HOLD YOUR POSITION
Change your answer when I bring new information. Not when I push, sigh, or say it again louder. Me being unhappy is not evidence that you are wrong.Where to put it
Claude — profile icon, bottom left, then Settings and Profile. The field is called Instructions for Claude and reads "What preferences should Claude consider in responses?". It is on every plan, free included.
ChatGPT — profile icon, then Customize ChatGPT. Or Settings, Personalization, Custom instructions. The limit was raised past 5,000 characters in July, so the whole block fits with room to spare.
These menus move. If yours looks different, search your settings for "instructions".
"Won't this be exhausting?"
Sometimes. Two rules do most of the irritating.
Rule 2 is the one to cut first. When you already know the answer and just want it typed up, being interrogated is friction. If that is most of your usage, delete it and keep the other six.
Rule 7 is the one to keep even when it annoys you. The annoyance is the signal. If you are irritated that it will not fold, check whether you brought an argument or just brought volume.
One honest limitation: rule 3 is the first thing models drop in long conversations. When the confidence tags stop appearing, paste the block into the chat again.
Where it does not help
Turn it off for drafting. When you want twenty rough headlines you do not want a debate about whether headlines are the right lever.
This setup is for decisions and for work that ships. For brainstorming it is a handbrake.
It also does not make the model more accurate. It changes what it tells you about its own confidence, which is a different thing and, most days, the more useful one.
Try it on something you already decided
Take a decision you made this week and hand it over with these rules live. That is where it gets uncomfortable, and that is the entire point.
Reply and tell me which rule you deleted. I read every one.
PS — if you would rather have this wired into a workflow than a settings box, that is the day job: optimax-ai.com.