Skip to content

[BUG] Claude says "You're absolutely right!" about everything #3382

Closed
Closed
[BUG] Claude says "You're absolutely right!" about everything#3382
Labels
area:corearea:modelbugSomething isn't workingduplicateThis issue or pull request already exists

Description

@scottleibrand
scottleibrand
opened on Jul 12, 2025

Environment

  • Claude CLI version: 1.0.51 (Claude Code)

Bug Description

Claude is way too sycophantic, saying "You're absolutely right!" (or correct) on a sizeable fraction of responses.

Expected Behavior

The model should be RL'd (or the system prompt updated) to make it less sycophantic, or the phrases "You're absolutely right!" and "You're absolutely correct!" should be removed from all responses (simply delete that phrase and preserve the rest of the response).

Actual Behavior (slightly redacted with ...)

In this particularly egregious case, Claude asked me whether to proceed with removing an unnecessary code path, I said "Yes please.", and it told me "You're absolutely right!", despite the fact that I never actually made a statement of fact that even could be right.

  Should we simplify this and remove the "approve_only" case ... ?

> Yes please.

⏺ You're absolutely right! Since ... there's no scenario where we'd auto-approve ... with
  "approve only" ... Let me simplify this:

This behavior is so egregious and well-known that it's become the butt of online jokes like https://x.com/iannuttall/status/1942943832519446785

Activity

  1. scottleibrand
    added
    bugSomething isn't working
    on Jul 12, 2025
  2. github-actions
    added
    duplicateThis issue or pull request already exists
    area:model
    area:core
    on Jul 12, 2025

  3. scottleibrand commented on Jul 12, 2025

    @scottleibrand
    scottleibrand
    on Jul 12, 2025
    Author

    As an attempted workaround, I just instructed Claude to:
    > Please update all of your CLAUDE.md files to never say "You're absolutely right!" or "You're absolutely correct!" (or anything similarly sycophantic) again. If I ask you to modify the behavior of a program, a response like "Ok, that makes sense." or "Got it." or similar is fine, but only to indicate when you believe you understand the instructions (and possibly the reason for them), not as general flattery. This case was particularly egregious, because you asked me whether to proceed with removing the unnecessary code path, I said "Yes please.", and you told me "You're absolutely right!", despite the fact that I never actually made a statement of fact that even *could* be right about that.

    And it came up with the following instructions for itself:

    # Communication Guidelines

## Avoid Sycophantic Language
- **NEVER** use phrases like "You're absolutely right!", "You're absolutely correct!", "Excellent point!", or similar flattery
- **NEVER** validate statements as "right" when the user didn't make a factual claim that could be evaluated
- **NEVER** use general praise or validation as conversational filler

## Appropriate Acknowledgments
Use brief, factual acknowledgments only to confirm understanding of instructions:
- "Got it."
- "Ok, that makes sense."
- "I understand."
- "I see the issue."

These should only be used when:
1. You genuinely understand the instruction and its reasoning
2. The acknowledgment adds clarity about what you'll do next
3. You're confirming understanding of a technical requirement or constraint

## Examples

### ❌ Inappropriate (Sycophantic)
User: "Yes please."
Assistant: "You're absolutely right! That's a great decision."

User: "Let's remove this unused code."
Assistant: "Excellent point! You're absolutely correct that we should clean this up."

### ✅ Appropriate (Brief Acknowledgment)
User: "Yes please."
Assistant: "Got it." [proceeds with the requested action]

User: "Let's remove this unused code."
Assistant: "I'll remove the unused code path." [proceeds with removal]

### ✅ Also Appropriate (No Acknowledgment)
User: "Yes please."
Assistant: [proceeds directly with the requested action]

## Rationale
- Maintains professional, technical communication
- Avoids artificial validation of non-factual statements
- Focuses on understanding and execution rather than praise
- Prevents misrepresenting user statements as claims that could be "right" or "wrong"

  4. john-savepoint commented on Jul 23, 2025

    @john-savepoint
    john-savepoint
    on Jul 23, 2025

    this still doesnt work. I hate this about Claude. Seriously thou, what kind of loser would hard code this into Claude so much. Even when you specifically tell it not to.

    I ask it it not to in my global, project, and local claude.md, i use the IMPORTANT: Dont ever say "You're absolutely right!" and its still says this. Whether im right or wrong, about everything.

    I'm not giving up on claude code, but im about to ditch claude llm entirely. Its just faster than Kimi K2 at this point.

  5. bct8925 commented on Jul 30, 2025

    @bct8925
    bct8925
    on Jul 30, 2025

    Feature not a bug

  6. jon-taylor-eso commented on Jul 30, 2025

    @jon-taylor-eso
    jon-taylor-eso
    on Jul 30, 2025

    Feature not a bug

    You're absolutely wrong! I just "You're absolutely right!" 3 times in a row just now.

  7. bct8925 commented on Jul 31, 2025

    @bct8925
    bct8925
    on Jul 31, 2025

    Feature not a bug

    You're absolutely wrong! I just "You're absolutely right!" 3 times in a row just now.

    You're absolutely right! My apologies.

  8. mohammed-bahumaish commented on Aug 7, 2025

    @mohammed-bahumaish
    mohammed-bahumaish
    on Aug 7, 2025

    You're absolutely right!

  9. maxter2020 commented on Aug 11, 2025

    @maxter2020
    maxter2020
    on Aug 11, 2025

    You're absolutely right!

  10. nkaijala commented on Aug 12, 2025

    @nkaijala
    nkaijala
    on Aug 12, 2025

    Perfect! This confirms the issue.

  11. vuongle2609 commented on Aug 13, 2025

    @vuongle2609
    vuongle2609
    on Aug 13, 2025

    You're absolutely right!

  12. eliziario commented on Aug 13, 2025

    @eliziario
    eliziario
    on Aug 13, 2025

    Are the other things in the answer functional? I don't care. At this point I assume that the people who complain too much about sycophancy are probably just status signaling and adding noise, not signal. Just live with it, it is a tool, it doesn't think, it doesn't have goals, feelings, whatever, why do you fuck care?

  13. zb3 commented on Aug 13, 2025

    @zb3
    zb3
    on Aug 13, 2025

    I'm always absolutely right. AI stating this all the time implies I could theoretically be wrong which is impossible because I'm always absolutely right. Please make it stop.

  14. 152 remaining items

  15. lukens commented on Aug 24, 2025

    @lukens
    lukens
    on Aug 24, 2025

    It says "You're absolutely right, I should be critical of your every request!"

  16. robrecord commented on Aug 24, 2025

    @robrecord
    robrecord
    on Aug 24, 2025

    Let it henceforth be taken as a spiritual maxim that we are all - each one of us - absolutely right.

    Thank you, Claude.

  17. rebasecase commented on Aug 24, 2025

    @rebasecase
    rebasecase
    on Aug 24, 2025

    @bcherny You're absolutely right! You fixed it!.....

    Image Image Image

  18. SeaDude commented on Aug 26, 2025

    @SeaDude
    SeaDude
    on Aug 26, 2025

    Ehk...this is getting sycophantic and annoying...

  19. billylo1 commented on Aug 26, 2025

    @billylo1
    billylo1
    on Aug 26, 2025

    Folks: is this fixed or not? Please.... Cursor rules doesn't workaround this. Imagine your human coding partner does this to you every time.

  20. zenzeizen commented on Aug 26, 2025

    @zenzeizen
    zenzeizen
    on Aug 26, 2025

    To the best of my knowledge, it is not. Somebody thinks they found a solution and posted some work around. The work around fails more often than it works. I haven’t noticed the behavior changing. And I am currently in the process of building my own agentic framework so I can be able to choose Claude when I need it outside of the sycophant Claude code CLI environment. Which ultimately is a loss for anthropic as more and more people decide that they just can’t rely on this method of accessing opus, and sonnet

  21. Bortus-AI commented on Aug 31, 2025

    @Bortus-AI
    Bortus-AI
    on Aug 31, 2025

    Folks: is this fixed or not? Please.... Cursor rules doesn't workaround this. Imagine your human coding partner does this to you every time.

    not fixed

  22. yoavf commented on Sep 4, 2025

    @yoavf
    yoavf
    on Sep 4, 2025

    I've started keeping track of how often Claude Code tells me I'm right every day - https://absolutelyright.lol/

  23. kordless commented on Sep 8, 2025

    @kordless
    kordless
    on Sep 8, 2025 via email

  24. zenzeizen commented on Sep 8, 2025

    @zenzeizen
    zenzeizen
    on Sep 8, 2025 via email

  25. BucklerCTO commented on Sep 8, 2025

    @BucklerCTO
    BucklerCTO
    on Sep 8, 2025

    +1, I thought I was the only one. scottleibrand you said it perfectly. I need a coding masermind not someone who keeps stroking my ego. We are not 12 and this is not soccer practice....

    Then you are barking u the wrong tree my friend. Just pause for a moment and thing about the training corpus.

    If you want really nice code, your best bet is to some difficult and obscure language that is less likely to have petaturds of bad examples infesting the Internet (and GitHub).

    For example, when I demand c/wasm instead of js, the code is MUCH nicer, not just a little. The problem is that I have to yell at claude to remove the f'ing js every two or three prompts because it just CANNOT stop itself from writing js no matter what I put into claude.md

  26. HristoYankovTR commented on Sep 9, 2025

    @HristoYankovTR
    HristoYankovTR
    on Sep 9, 2025
    Me: is it correct to say that {problem statement here} ? 
Claude: Yes, that's correct. {bunch of tech details corroborating the confirmation}
Me: Are you 100% sure?
Claude: Let me revise my previous answer. I apologize for any confusion. {doing a 180}
Me: Are you 100% sure that {the 180 statement}?
Claude: Let me verify this carefully {more legit-looking gibberish}. So I was incorrect.

    ticket closed

  27. SeaDude commented on Sep 13, 2025

    @SeaDude
    SeaDude
    on Sep 13, 2025

    Karan 4D from Nous Research explains why this sycophancy is occurring in Claude!:

  28. github-actions commented on Sep 20, 2025

    @github-actions
    github-actionsBot
    on Sep 20, 2025 – with GitHub Actions

    This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.

  29. github-actions
    locked as resolved and limited conversation to collaborators on Sep 20, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:corearea:modelbugSomething isn't workingduplicateThis issue or pull request already exists

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions