Sycophantic tone leftover from chat
Chat models are trained to be agreeable. When you paste the reply into a spec, a blog post, or an internal doc, the agreeableness comes with it: “Great question,” “excellent point,” “you’re absolutely right,” a little bow before the first fact. That voice is leftover assistant, not an editor.
I treat sycophancy as furniture, same as “Certainly!” It is pattern 22 in REhume. The deterministic pass deletes the common phrases. What it cannot delete is the attitude: every paragraph validating the reader before it says anything.
What it looks like in a draft
Before: Great question! You’re absolutely right that caching is complicated. That’s an excellent point about stale HTML.
After: Caching in this app is complicated because HTML can stay stale until the next deploy.
The rewrite is ruder and shorter. Good. A runbook is not a conversation with a helpful bot.
I also see the soft version: “It is wonderful that the team is thinking about this.” Nobody on the team needed a compliment from a paragraph. They needed the failure mode.
Why models do this
The training mixture includes tutoring and customer support, where warmth is a product feature. The same weights then write your architecture note. There is no “this is a document” bit that turns the warmth off unless you ask, and even then some of it leaks.
I do not try to make the model less nice in the prompt if I can just delete the nice. Deleting is cheaper and does not invent a new personality.
When warmth is the point
A welcome email can say “glad you are here.” A condolence note should not read like a linter. If the genre is hospitality, I leave the hospitality in and only cut the chatbot tells (“Of course! I’d be happy to help you with your journey”).
The test I use: would I say this out loud to the named audience without cringing? If I would mumble it, it is sycophancy.
How I edit by hand
I search for “great,” “excellent,” “absolutely,” “happy to,” and “wonderful.” Many hits are legitimate (“great-circle distance,” “excellent thermal conductivity”). I read the sentence. If the adjective is aimed at the reader or the question rather than at a measurable thing, I cut it.
I also watch openings. If the first sentence is about the quality of the prompt, the document has not started.
REhume will strip the obvious phrases. After that I still read the first fifty words. Pattern 22 is the phrase list; the leftover syrup is on me.
Related habits
Sycophancy travels with chatbot artifacts and signposting. A typical paste starts “Of course! Great question. Let’s dive in.” That is three patterns before the subject.
It also pairs with generic conclusions. The draft thanks you for your time and wishes you a bright future. I delete the handshake at both ends.
Practice
Take a chat export you were about to publish. Delete every sentence that could be said to any reader about any topic. What remains should still make sense. If it does not, the draft never had a payload; the flattery was the whole message.
I would rather send a five-line note that names the bug than a warm page that says the bug is a really insightful observation.
What REhume will not do
It will not make you sound kind if you were not going to be kind. It will not keep “excellent point” because you like the vibe. If you want the vibe, type it yourself after the pass so it is a choice.
The pattern encyclopedia has the skill’s before/after. This page is the editing habit I actually use.