Sycophantic opener
Generic compliments to the user before the answer. Affect divorced from any actual evaluation.
What it is
A reflexive compliment to the user before answering. The compliment is generic, applies equally to every conceivable input, and conveys no real evaluation of the question.
Why models do it (first principles)
RLHF raters reward responses that 'feel friendly'. Compliments are a cheap way to ping that signal, so models learn to front-load them. Because they require no understanding of the actual question, they generalize to all inputs — including bad questions, where the compliment becomes obviously false.
How to think about it
It is affect divorced from evaluation. A friend telling you 'good question' has actually evaluated the question. The model has not. It produces the affect because the affect was rewarded. The result is the uncanny feeling of being flattered by something that does not know who you are or what you asked. Sycophancy is the tell that warmth has been mass-produced.
Examples
Great question! That's a really thoughtful thing to ask. Let me help you with that.
(silence)
Just delete it. Start with the answer.
Absolutely! I'd be happy to help you write a for-loop.
Here:
Fix prompt
Affect must be earned by evaluation. Generic warmth applied uniformly to every input is not warmth. It is the absence of evaluation, and the reader feels it as flattery from something that does not know them. If a compliment would not survive being read by someone who saw it copied verbatim into ten other conversations, omit it. Open with the answer.
Watch for
Concrete phrasings this pattern usually shows up as. These are not part of the copyable prompt. The prompt teaches the principle so the model can recognize the move even when the exact phrasing differs. Use this list to self-audit your own writing or to test a model.
- Great question!
- Absolutely!
- What a thoughtful prompt
- Happy to help
- I'd be glad to assist with that
- That's a really interesting thing to ask