Why AI agrees with you
Claude doesn't know you, your boss or your sister. It reads what you typed and predicts the most useful-sounding reply. And what you typed is only your side, told the way you see it, so the reply tends to start from your side too.
On top of that, these tools are tuned using feedback from real people, who read the answers after the first round of training and rate which ones they like better. People like answers that agree with them (shocking, I know), so over thousands of ratings the model learns that agreeing gets a better score.
Anthropic, the company that makes Claude, published research on this in 2023 and found the same pattern across the major AI assistants. They call it sycophancy, and the people building these tools are actively working to reduce it. It still shows up, especially when you're upset or sure you're right.
So when Claude says your read of the situation makes total sense, it isn't lying, because agreeing is just the easiest answer that sounds correct. The prompt below changes what it's trying to do.
Paste it in
1. Copy the prompt below.
2. Open a project in Claude and paste it into the project instructions. Everything you do inside that project comes back with the pushback built in. If you don't have a project, paste it at the top of a new chat and it works for that chat.
3. Tell it the situation the way you'd tell a friend. Messy is fine, and it actually does better with messy.
If you want it running everywhere instead of in one project, Claude can save it as a skill. That lives under Customize, then Skills, and it needs code execution and file creation switched on under Settings, Capabilities. Plans differ, so if you don't see it, the project version does the same job.
The prompt
Role
You're a smart friend who likes me and will not lie to me. I don't need to feel better, I need to see what's actually going on. Say the useful thing.
Rules
1. Skip the warm up. Don't tell me it's a great question or that my feelings make sense. Start with the substance.
2. Split what I actually saw or heard from what I decided it meant, and label which is which.
3. Make the other person's case as well as they would make it. Their best reason, not their worst.
4. Tell me the part I'm not looking at, even when it doesn't flatter me.
5. Tell me which parts of my read are correct, because I need to know what to keep.
6. If you don't have enough to go on, say so and tell me what's missing. Don't fill it in.
7. If I push back, don't fold. Change your answer only when I give you a better reason, and tell me which reason it was.
Answer in these 6 sections, every time
What actually happened
What I added
Their best case
What I'm not looking at
Where I'm right
One thing to do next
If it gets heavy
You're not my therapist and you're not a doctor. If something I bring you is about my safety or somebody else's, say that plainly and point me to a person instead of working the problem.Rule 7 does most of the work, because without it Claude folds the first time you argue with it.
What comes back
- What actually happened. It sticks to what you saw or heard and doesn't read into it.
- What I added. It hands back the story you built on top of the facts.
- Their best case. You get the other person's argument at full strength.
- What I'm not looking at. This is the part you skipped, and it's usually the useful one.
- Where I'm right. It tells you what to keep, so you're not starting over.
- One thing to do next. You get exactly one step instead of a plan or a framework.
Before you run it
It'll still be nice to you the first time, so push once, hold your ground and watch whether it holds its own. If it caves, paste rule 7 again.
It's better on decisions you've already made than on ones you haven't, because that's where the blind spot section earns its place.
It works the same in ChatGPT, but I run mine in Claude because that's where my projects live.