⚠️ Stop telling AI it's wrong

⚠️ Stop telling AI it's wrong. That's the fastest way to turn a correct answer into a worse one. Anthropic found that models flip accurate answers and apologize for mistakes they never made when you push back. These are the 7 follow-ups you can use to fix a bad answer instead of making it cave: 1. When it caves the moment you doubt it. Pushing back reads as a signal to agree with you, so drop the signal. Say: "I'm not saying you're wrong. Re-examine your answer on its own merits. If it holds up, defend it and tell me why. If it doesn't, show me the exact mistake." 2. When it states made-up facts with full confidence. Make it rate its own certainty before you trust a word. Say: "Go through your answer claim by claim. Label each one confident, inferred, or unsure. For anything not confident, tell me what would verify it." 3. When it answers a different question than you asked. Get it to repeat the task back before it burns another response. Say: "Before you redo anything, tell me in one sentence what you think I'm asking for. If any part is unclear, ask me instead of guessing." 4. When it ignores a rule you gave it. Force a line-by-line check against your own instructions. Say: "List every requirement I gave you. Go through your answer and mark each one met or missed, with the line that proves it." 5. When a long answer contains a small error. Tell it an error exists so it hunts instead of skims. Say: "There is at least one error in the reasoning above. Find it and fix it. Walk through each step until you catch where it breaks." 6. When the answer is correct but shallow. Raise the bar and make it show you the gap. Say: "That's a beginner-level answer. Redo it as a top expert in this field would, and list what they'd include that you left out." 7. When it keeps building on a wrong approach. Cut the anchor and name the flaw so it doesn't repeat. Say: "Drop your previous approach completely. It failed because [reason]. Start fresh and don't reuse that approach or its conclusions." Every one of these gives the model a job instead of a verdict. "You're wrong" tells it to flip. "Find the error in step 3" tells it to think. The people who get clean answers out of AI direct the model. They don't argue with it. @aipost 🏴