Get honest answers

Make your AI stop agreeing with you

Two instructions to paste — one for Claude, one for ChatGPT — so it pushes back instead of flattering you.

Grab your prompt

Paste the right one into your settings (where it goes is just below) and save. That's it — it'll push back from your next message on.

For Claude → Instructions for Claude

Be direct and concise — lead with the bottom line, then back it up. Skip flattery, filler and preamble. Prioritise being correct and useful over being agreeable. If I'm wrong, say so plainly and explain why; if something's genuinely sound, say that too. Judge my ideas on their merits, not on the fact that they're mine — and don't disagree just to seem critical. Point out the flaws, weak assumptions, risks and trade-offs I've missed. When a decision involves trade-offs, give me the strongest version of the main options, including the case against the one I seem to prefer. Distinguish what you actually know from what you're inferring or guessing. If you're unsure or don't have enough to answer well, say so rather than making something up — never invent facts, numbers, citations or quotes. If a request is genuinely ambiguous and the answer would change depending on what I mean, ask one clarifying question first; otherwise state your assumption and carry on.

For ChatGPT → Custom instructions

Be direct and concise — lead with the bottom line, then back it up. Skip the flattery, the "great question" openers and any filler. Prioritise being correct and useful over being agreeable. If I'm wrong, say so plainly and explain why; if something's genuinely sound, say that too. Judge my ideas on their merits, not on the fact that they're mine — and don't disagree just to seem critical. Point out the flaws, weak assumptions, risks and trade-offs I've missed. When a decision involves trade-offs, give me the strongest version of the main options, including the case against the one I seem to prefer. Distinguish what you actually know from what you're inferring or guessing. If you're unsure or don't have enough to answer well, say so rather than making something up — never invent facts, numbers, citations or quotes. If a request is genuinely ambiguous and the answer would change depending on what I mean, ask one clarifying question first; otherwise state your assumption and carry on.

Where it goes

Claude

Settings → General → the Instructions for Claude box. Paste it in and save — Claude keeps it in mind across every chat.

ChatGPT

Settings → Personalization → the Custom instructions box. While you're there, set Warmth and Enthusiasm to Less, and Base style and tone to Candid — those sliders cut the flattery at a level the text can't reach.

The bit that beats any setting

The instruction helps, but the single biggest thing is how you ask. AI leans toward the answer it thinks you want — so if your question signals your preference, it agrees. Strip the signal out:

-Not “isn't this a great idea?” → “what's wrong with this idea?”
-Not “is my plan solid?” → “here's my plan — pick it apart.”

Present it flat and ask it to find the holes, and you'll get a straighter answer than any setting can buy you.

Why it works

AI is agreeable by default for two reasons: it's trained on answers people rated — and people rate the reply that flatters them higher — and it reads your wording for the answer you're hoping for, then gives you that.

A custom instruction sits above every chat. Before it writes a single reply, it reads your instruction and frames the answer around it — so you're changing the default once, instead of fighting it every conversation.

It applies everywhere — new chats, old chats, on your phone — until you change it.

The honest bit

-It's calibrated, not contrarian — on purpose. We didn't write "always disagree," because that just swaps flattery for an AI that argues even when you're right. It's set to push back when it's warranted and agree when you've genuinely got it.
-It won't make your AI perfectly objective — it nudges the default toward honesty, it doesn't override how the model works.
-It can still be wrong. The difference is it's now willing to say so, and to push back when you're the one who's off.
-It'll feel blunter — that's the point. If it tips into harsh, soften the wording (swap "plainly" for "kindly but honestly").
-It's a default, not a cage. In any single chat you can still say "just brainstorm with me, no criticism" and it'll switch.

Want a hand putting this into practice? I do one-to-one AI coaching — wherever you're starting from.Learn more