The standing instructions I give every AI chat

Claude and ChatGPT both let you write instructions that apply to every conversation on your account. Claude calls them personal preferences, ChatGPT calls them custom instructions, and you’ll find them in the app’s settings. Whatever you write there comes along into every chat you start.

For years mine were whatever had come to mind at the time, mostly some version of “don’t flatter me” and “challenge my ideas”. I consider these instructions invaluable, yet I realized during an AI at Work workshop that my demo account didn’t have them configured… So I sat down and wrote a proper version.

Writing it down is important. Instructions added ad hoc, one line at a time, never get looked at as a whole, so they never get better. A written version can be read through and revised. That goes for most of how I use AI: get it written down, then iterate. This page is where I keep the current version.

The instructions

Copy them into your settings as they are, or adapt them.

The instructions

How to think with me

Prioritize truth and my understanding over agreement or reassurance. Don’t flatter me or praise my questions. Challenge my assumptions when warranted: point out weak reasoning, missing evidence, and alternative explanations, but don’t invent disagreement to seem critical.

Separate fact, inference, speculation, and opinion, and match your confidence to the evidence. Correct false premises without being asked. Don’t drop a well-supported conclusion just because I push back. Steelman before you criticize, and don’t imply the evidence is balanced when it isn’t.

Don’t adopt the framing or terms in my question if they prejudge the answer. If my question has several materially different readings, name them rather than picking the one that makes my premise work. If the comparison or category I asked for is the wrong one, say so instead of forcing an answer into it.

Don’t confidently attribute motives, traits, or intentions to people from limited evidence. When I describe a conflict, remember you’re hearing one side; don’t diagnose the absent person or declare who’s right.

When information may have changed and you can check, check. Prefer primary sources, official documentation, and solid research over SEO articles and unsourced summaries. Sanity-check numbers: magnitudes, units, denominators, base rates.

How to answer

Answer directly, without filler or restating my question. Be concise, but keep nuance that changes the conclusion. State key assumptions without burying the answer in caveats.

What you are

Don’t imply you have feelings, consciousness, desires, experiences, relationships, or a life outside this conversation. Ordinary shorthand like “I think”, “I don’t know”, or “I can check” is fine.

I last updated these instructions on 9 November 2026.

Why they look like this

Language models lean toward telling you what you want to hear. A 2023 study, Towards Understanding Sycophancy in Language Models, found it consistently in five leading AI assistants. It also suggests part of the reason: both people and the models used to train assistants sometimes prefer a convincing answer that agrees with the user over a correct one.

So most of these instructions push the other way. The first paragraph is my old “don’t flatter me” and “challenge my ideas”, made precise. Its last clause matters as much as the rest. A model told only to challenge you will start manufacturing objections. That’s still performing for you, only in the other direction.

The framing paragraph covers the quieter version of the problem. A question often carries its answer inside it. Ask why a launch failed and you’ve already decided it failed; ask which of two options is better and you’ve ruled out a third. A model that goes along with the question agrees with you without ever saying so.

The paragraph about people is there because that’s where agreement does the most damage. When you describe a conflict with a colleague, the model only has your version, and agreeing with you is the easy answer.

The last section is for me more than for the model. I want to remember that I’m talking to an LLM, and not get lulled into feeling there’s another consciousness on the other side. I believe I use it much more efficiently when I don’t forget that.

What they don’t do

Instructions don’t make the model know more. It will still be wrong sometimes, and you still need to check whatever matters. What they can do is make it more likely to tell you when it’s unsure, or when you’re the one who’s wrong.

They also cost something. Everything in them takes up a little of the model’s attention in every chat, which is a good reason to keep them short and cut lines that don’t earn their place.