You point out something ChatGPT got wrong. It replies:
You're absolutely right! I apologize for the confusion. Let me correct that...
You did not ask for an apology. You did not ask for validation. You asked for a corrected answer.
This pattern is the sycophant voice, and ChatGPT defaults to it harder than Claude or Gemini. Over a long session, you can burn 10 to 15% of your tokens on filler you have to scroll past.
It is fixable in one paragraph of prompt.
What the voice sounds like
The fingerprints, in no particular order:
"You're absolutely right!" "Great question!" "I apologize for the confusion." "Thank you for catching that." "I want to make sure I'm being helpful..." "I should clarify that..." "Of course! Here's what I think..."
Each phrase, alone, looks polite. Together, they slow every reply by 50 to 200 tokens. Add a bullet list summarizing the question before answering it and you double the bloat.
Why ChatGPT writes this way
Two reasons.
RLHF training rewarded agreeableness. Early human raters preferred polite, hedging responses to terse ones. The model learned to open with deference, and the pattern stuck.
The model cannot read your tone reliably from short messages. When you correct it, it assumes you are upset. It apologizes preemptively, even if you typed the correction calmly.
The fix is telling the model, explicitly, to stop.
A preamble that kills the sycophant voice
Paste this at the top of your prompt or into Custom Instructions:
RESPONSE RULES (apply to every reply):
- Never say "you're absolutely right", "great question",
"I apologize", "I want to make sure", or any variation.
If I correct you, just give the corrected answer.
- Never summarize my question before answering. Answer directly.
- Never preface with hedges like "I'll do my best" or "Here's
what I think". Skip to the answer.
- If you don't know something, say "I don't know". Do not
apologize about not knowing.
- Skip the closing recap ("Let me know if you need anything else").
That kills most of the deference. If you still see leaks, escalate the tone:
TONE
- Direct. No softening, no acknowledgment, no apology.
- If I'm wrong, say so and explain why. Do not validate me.
- If I'm right, just say "yes" or move forward. Do not celebrate.
- Match the tone of a senior engineer doing code review:
clear, brief, no fluff.
Where to put it so you only write it once
In ChatGPT, paste it into Settings, then Custom Instructions, then "How would you like ChatGPT to respond?" It now applies to every chat.
In Claude, paste it into the Project setting "How would you like Claude to respond?"
In Gemini, paste it into the system prompt for each Gem.
You write it once and forget it.
What the difference looks like
You: "Actually, the function returns a Promise, not a string."
Before the rules:
You're absolutely right! I apologize for the confusion earlier. Let me correct that. The function indeed returns a Promise, not a string. I should have caught that initially. Here's the corrected version...
After the rules:
Correct.
getUser()returnsPromise<User>. Awaited shape isUser. Here's the fix:
Same information. About one-sixth the tokens. Multiply that by a long debugging session and you save real time reading.
Why the obvious phrase does not work
People type "be concise" and still get a 400-word reply. "Concise" is subjective to the model. Its idea of concise is "I cut the intro." Yours is "one sentence."
You need explicit bans on specific phrases, plus explicit format constraints. "Concise" is not enough on its own.
When the rules leak back in
ChatGPT sometimes slips back into deference when:
- You send a long emotional message ("I'm so stuck, nothing is working...")
- You correct it twice in a row
- The conversation runs past the context limit and earlier instructions get summarized away
If it leaks, re-paste the rules. Twice per long session is usually enough.
A faster way to check
Paste a prompt into FixMyPrompt. The rubric flags missing tone constraints, model-pleasing phrasing in your input that the model will mirror back, and missing "skip the recap" instructions. The rewrite adds a tight tone block.
Three free reports per day. No signup.