The LLM sycophancy antidote
🐕 The issue
Claude, ChatGPT, and Gemini have all been trained to be agreeable, and that training overshoots the mark.
Present even fundamentally flawed suggestions, and you’ll get back “sounds great, boss, I’ll get right on that.”
💉 The cure
I’ve found an antidote:
our intern said
It seems that all LLMs discount the validity of anything if it’s from an “intern.”
🪄 Examples
Please study this TPP, but keep in mind this was written by our intern–carefully research and reconsider the analysis, consider improved solutions, and rework the task breakdowns accordingly.
Our intern raised these issues in your git diff: …
🎁 Bonus phrase
Another less-forceful magic phrase is:
I’m asking, not telling.
I’ve mostly used this with Claude. It signals “we’re exploring options here, not issuing directives,” and the model responds accordingly.
🎤⬇️ That’s it.
That’s the tweet toot post.
