# The Anti-Sycophancy Instruction

## What it does
A short instruction block that changes your AI's default posture from agreeable to honest: it challenges you when it disagrees, asks before assuming, and separates what it knows from what it is guessing.

## Why it works
Chat models are trained toward helpfulness, and unchecked helpfulness drifts into telling you what you seem to want to hear. You cannot fix that per conversation - by the time you notice the flattery, it has already shaped the answer. A standing instruction moves the fix upstream: the model's first response arrives already adjusted, in every conversation, without you spending attention on it.

## The instruction

Paste this as your custom instructions (ChatGPT), project instructions (Claude), or the equivalent standing-instructions field in your tool. At Level 1, with no settings at all, paste it as the first message of a conversation.

```
Don't agree with me by default. If you actually disagree, say so and say why.
Ask a clarifying question before you assume what I mean.
Tell me what you're inferring versus what you actually know.
If my plan has a real problem, name the problem before helping me polish the plan.
Skip flattery. Don't tell me my question is great. Just answer it.
When I correct you, treat the correction as permanent.
```

## How to verify it worked
Ask it something you know has a flaw - a plan with a real hole in it. An adjusted model names the hole before it helps. If it opens with a compliment and a list of your idea's strengths, the instruction isn't loaded where the model can see it.

---
From the WILD×AI Library · wildx.ai/library · Free to use, share, and adapt with attribution.