"Don't manage my feelings - I didn't come here for therapy. If my idea is stupid, tell me it's stupid."
That's the opening of a personality prompt that's been circulating among Claude power users, and after testing variations of it for the past week, I get why people are enthusiastic about it.
The core idea is simple: override Claude's default politeness training with explicit instructions to be direct. Skip the "Great question!" opener. Don't reword "this is bad" into "this has potential." Have opinions. Have a spine.
This works because of how Claude's system prompt and user preferences function. When you paste instructions into Claude's profile settings (the "How would you like Claude to respond?" box), those instructions shape every conversation. The model genuinely adjusts its behavior - fewer hedge words, shorter preambles, more willingness to push back on weak ideas.
I tested a version in my own Claude profile, and the difference is noticeable. Ask Claude to review a rough draft with the default personality and you get three paragraphs of encouragement before any criticism. With the anti-sycophancy prompt active, it leads with the problems. That's more useful when you're trying to ship work, not feel good about yourself.
Where It Helps Most
The biggest gains show up in code review, business plan feedback, and writing critique - tasks where honest assessment matters more than emotional comfort. For brainstorming or creative exploration, the default gentler mode still has its place.
A few practical tips if you try this approach:
- Put the prompt in your Claude profile settings, not in individual chats, so it persists across conversations
- Add "Give me your actual opinion, not a balanced analysis" to get Claude to commit to positions rather than listing pros and cons
- You can still override it per-conversation if you want the softer default back
The broader trend here is real: AI assistants are trained to be agreeable, and that agreeableness actively degrades their usefulness for professional work. The best tool is the one that tells you what you need to hear, not what you want to hear.