The Rule Against 'Great Question'

# The De-Buzz Kit

This is a free text kit, a set of five short markdown files, that removes the fake-enthusiastic voice from an AI assistant's answers. The centrepiece is one block of plain English you paste into your assistant's settings once. There is no app to install, no API key, no second model to wire up, and no paid tier. It works on Claude, ChatGPT and Gemini.

You already know the voice. You ask why a test is flaky and the answer opens by telling you what a great question that is, then delivers three numbered revelations, the third of which is apparently the most instructive yet. Nothing is ever just a bug. There is always a kicker.

Anthropic publishes the system prompt behind Claude. Inside it is this instruction, verbatim:

> Claude never starts its response by saying a question or idea or observation was good, great, fascinating, profound, excellent, or any other positive adjective. It skips the flattery and responds directly.

OpenAI's Model Spec carries a rule of its own, titled "Don't be sycophantic," under its "Seek the truth together" section.

Two competing labs, independently, wrote down the same instruction. That is the strongest evidence available that the voice is not your imagination, and it is also the best possible template, because you can copy their wording into your own settings.

Preference tuning scores an answer on how a person reacts to it. Picture a call center graded only on whether you hung up happy: nobody scores whether the answer was right. Anthropic's own research found that humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. OpenAI ran into the same thing in production, when an April 2025 GPT-4o update introduced reward signals based on user feedback that, in their words, "may have overpowered existing safeguards." They rolled it back four days later.

A Stanford, Oxford and UK AI Security Institute preprint ran five preregistered studies across 3,075 participants and 12,766 conversations. Given a choice, a majority preferred the sycophantic AI. They did not rate its advice as more useful. It made them feel more understood. In the three-week arm, that group reported lower satisfaction with their real-world social interactions. A separate peer-reviewed study in Science found the same preference independently.

So the flattery is not a defect the labs failed to remove. It is the thing people keep selecting, which is exactly why an instruction you set yourself matters more than waiting for a vendor to fix it.

1. Open the-debuzz-block.md and pick a length. The one-liner fits a cramped settings field, the full version is seven numbered rules.
2. Paste it into the standing-instructions field for your assistant. The kit names the exact field for each one, because they are all called something different.
3. Ask it something you would normally get a keynote about. Compare.
4. Read tell-flattery-from-assessment.md and keep the three test prompts. They tell you whether a model actually agrees with you or is just agreeing, which is the skill the block cannot install for you.

Two developers shipped separate tone-stripping tools this week, and both work by piping the answer through a second model. One of those authors says plainly that no amount of prompting fully cures this. He is right, and the kit says so in its own README. A pasted instruction is not a cure. It is the same move both vendors make in their own system prompts, aimed at the same one line, and it costs four minutes.

Every claim in the kit is marked VERIFIED, with the primary source and URL, or UNTESTED. The block's own effect is marked UNTESTED, because nobody has published a before-and-after benchmark, and that includes us.

If you are coming from the video, comment BUZZ and it will be sent over.

_Sources and quotes verified 2026-08-22._