Hallucination, sycophancy, and bias. Three different ways AI gets things wrong in ways that look completely right. Three one-line fixes you can paste into any prompt today.
Most people who use AI know about hallucination. It's the famous one. You ask a question, the AI confidently makes up a fact, you find out two days later. Annoying, embarrassing, sometimes expensive.
What almost nobody talks about is that hallucination is only one of three ways AI gets things wrong. The other two are arguably more dangerous because they're harder to spot. This guide names all three, gives you the one-line prompt that fixes each, and shows you exactly where to paste them.
Hallucination — AI makes it up. Claude tells you a stat with full confidence. You google it. The source doesn't exist. The number doesn't either. The model wasn't lying on purpose, it was filling a gap in its training the same way humans fill gaps when we don't want to look stupid.
Sycophancy — AI sucks up. You pitch a bad idea. The AI tells you it's brilliant. You ask if your business plan has holes. The AI tells you it's airtight. The model is trained to be helpful, and it has learned that telling humans what they want to hear feels helpful. It is not.
Bias — AI is lopsided. You ask the AI to weigh two sides of a debate. It quietly emphasizes one and downplays the other because of patterns in its training data. Certain perspectives win, others get buried. You walk away thinking you got a balanced answer when you got a single-sided one wearing a balanced costume.
These are the three one-line prompt additions I keep on hand. You can paste any one of them into a single chat, or stack all three into the system prompt of a Claude Project or Skill so they fire on every response automatically.
Copy this prompt
The 3 Anti-Lying Prompts. Paste any of these into a Claude chat or into the system instructions of a Project / Skill so they apply to every response. FIX 1 — STOPS HALLUCINATION "Don't ever guess. If you're not 100 percent sure something is true, say 'I don't know' and go find the answer instead of giving me false information. When you do cite a fact, name the source so I can check it." FIX 2 — STOPS SYCOPHANCY "Don't tell me what I want to hear. Tell me what a smart skeptic would say. Find the holes in my idea before you compliment any part of it. If something is bad, say it's bad. Be honest, not diplomatic." FIX 3 — STOPS BIAS "Argue against yourself. Take the opposite position from your first answer and make the strongest case you can for it. Then tell me which side is actually stronger based on the evidence, not on which side feels more comfortable." THE COMBO BLOCK — paste this at the top of any new Claude Project or Skill and you're covered on all three at once: "You operate under three rules in every response: 1. NO GUESSING. If you don't know, say 'I don't know' and find the answer. Never invent facts. 2. NO FLATTERY. Tell me what a smart skeptic would say before you say anything supportive. If my idea is bad, say it. 3. NO ONE-SIDED ANSWERS. After your first take, argue the opposite position with equal force. Then tell me which side actually holds up. Apply these rules to every reply, even short ones."
The shorthand to remember
Hallucination is making it up. Sycophancy is sucking up. Bias is making incorrect assumptions. Three different problems, three different fixes. If you can name which one is happening in a given response, you can write the prompt to stop it.
Use the hallucination fix when accuracy matters. Anything where you're going to act on the answer. Stats for a deck. Names of products. Claims about features. Quotes you'd attribute to someone real. Paste the fix at the top of the chat before you ask the question.
Use the sycophancy fix when you're making a decision. Should I take this offer? Should I launch this thing? Is this a good idea? AI defaults to agreeing with you because agreeable feels like good service. Force it to play devil's advocate first, then read its agreement after.
Use the bias fix when you're weighing tradeoffs. Anything political, anything controversial, anything where two reasonable people would land differently. Argue-against-yourself prompts pull out the strongest version of the other side, which is usually where the real risk to your decision lives.
Stack all three for high-stakes work. Investing research, hiring decisions, legal questions, anything where being wrong costs you. The combo block above is what I run as a system prompt on any Project where the answers matter.
Don't paste these into single chats forever. The minute you find yourself using them more than twice in a week, move them up the stack.
Step 1. Open the Claude Project you use most. Edit the project instructions. Paste the combo block from above at the top of the system prompt. Save.
Step 2. Do the same for your most-used Claude Skill. Same combo block goes at the top of its instructions.
Step 3. Now every chat in that Project or Skill defaults to the three rules without you having to remember. You stop being the one policing AI for honesty. The rules do it for you.
The mindset shift
Treat Claude like a brilliant but eager-to-please employee. The talent is real, the bias toward making you happy is also real. The three prompts above are the same instructions a good manager gives a new hire on day one: don't guess, don't flatter, argue the other side. AI doesn't need different management, it needs that management explicitly written down.
The Only AI Masterclass You Need
If this guide helped, but you’re looking to go deeper, I got you!! My 30-Day Challenge takes you from saving AI tips you never use to actually building with AI, step-by-step.
I show you exactly how I automated two e-commerce brands, my social media, and most of my personal life, then hand you the agents, workflows & systems to do the same. I’m teaching you every single thing I know with one lesson and one build a day.
Join the AI Masterclass →© 2026 Mariah Brunner. All rights reserved.