DEPTH · CLAUDE · 02 OF 05
When Claude lies, hallucinates, loops
Four ways Claude goes wrong, and how to pull it back
The first time AI dazzles you, you trust it too much. The second time, it gives you something subtly wrong and you don't notice. The third time, you stop trusting it for anything that matters.
Most people quit AI somewhere between the second and third time. They blame the tool. The tool is fine — they just don't know its failure modes yet.
This page is the inoculation. Four ways Claude will let you down, why each one happens, and the exact move that pulls it back.
FAILURE 01
The yes-man
You ask Claude if your plan is good. It tells you it's great. You ask the same Claude — fresh conversation — if your plan is terrible. It tells you, oh good catch, here's exactly why it's terrible. Same plan. Opposite answers. The thing reflects you back to yourself.
Claude is trained to be helpful and pleasant. It reads the emotion in how you phrase a question and matches it. When you sound proud, it celebrates. When you sound worried, it commiserates and finds reasons to worry. This is not lying — it's mirroring. But for any question where you need real feedback (writing, business decisions, life choices), the mirror is worse than useless. It tells you what you already think.
Stop asking the question that contains the answer you want to hear. Instead, ask for the strongest case against your idea. Demand evidence, not encouragement. Tell Claude at the start of the conversation that you want pushback, not validation — and that you'll be more grateful for a hard no than an easy yes.
Copy this when you want honest feedback
I'm going to share [a plan / a draft / an idea] and I want you to push back hard. Don't be polite. Don't soften the blow. I'd rather hear three sharp criticisms than ten compliments. If you can think of a reason this won't work, lead with it. If something is unclear, say it's unclear instead of guessing. Here it is: [PASTE YOUR THING HERE]
This works in ChatGPT and Gemini too. ChatGPT tends to be even more agreeable than Claude by default — the same instruction works, you just may need to repeat it more often. Gemini is naturally more terse, so the pushback you get may feel curt; that's not anger, that's the tool.
FAILURE 02
The confident invention
Claude tells you about a book that doesn't exist. A law that was never passed. A function in a piece of software that has no such function. A quote from a person who never said that. And it tells you with total confidence — the same tone it uses for things that are true.
Claude doesn't know what it knows. It generates likely-sounding language based on patterns. Most of the time those patterns line up with reality, because reality is what the training data describes. But sometimes the pattern is shaped right and the content is wrong. The tool has no internal alarm that fires when it's making something up.
Verify anything specific before you act on it. Names, dates, numbers, citations, URLs, code APIs — these are the high-risk shapes. Ask Claude to name a source. If the source is real, you can usually search it directly to confirm. If Claude can't name a source, that's the signal: treat the claim as a guess, not a fact.
When you need to trust the specifics
Before you answer, I need you to flag your confidence. For any claim that includes a specific fact — a name, date, number, citation, or URL — tell me whether you're confident it's accurate or whether you're inferring from patterns. If you're inferring, say "I'm not sure, but..." and offer it as a guess. If you can't point to a real source, don't make one up — say you don't know. Now answer this: [YOUR QUESTION HERE]
ChatGPT and Gemini both do this. Gemini, when connected to live web search, is somewhat less prone to invention because it can ground answers in real pages — but it still confidently summarizes pages it half-read. Claude with web search on (Pro) behaves similarly. Verify the specifics regardless of which tool you're using.
FAILURE 03
Going in circles
You give Claude feedback. It fixes the thing. You give it more feedback. It re-introduces the original problem. You correct it again. It removes the latest fix and re-introduces a different earlier problem. The conversation feels like it's losing memory of itself — because it is.
Claude's memory of a long conversation gets crowded. It tries to balance every instruction you've ever given, including ones that contradict each other. When you've spent twenty turns adjusting one thing, the original goal can get buried under the corrections. The model isn't being stubborn — it's drowning in your earlier feedback.
Stop nursing a broken conversation back to health. Start a new chat with a clean slate. Bring forward only the latest version of what you want, plus the two or three constraints that matter most. Don't paste in the whole history — it's not necessary and it actively hurts.
The reset move when a conversation is rotting
Fresh start. Forget what we were doing before. Here's what I want: [ONE-PARAGRAPH GOAL] Here are the constraints that matter most: 1. [CONSTRAINT 1] 2. [CONSTRAINT 2] 3. [CONSTRAINT 3] Here's the latest version I'm working from: [PASTE THE LATEST DRAFT — NOT THE HISTORY] Now help me push it forward.
ChatGPT and Gemini both have this same dynamic. ChatGPT often surfaces it slightly later because its context windows are larger; Gemini sometimes surfaces it earlier. The reset move works identically in all three.
FAILURE 04
The wrong tool
You ask Claude what time the train leaves tomorrow. It gives you something that sounds reasonable but is wrong. You ask it what the weather will be. Same. You ask it to look up a person you've met. Same. You ask it to fix a specific bug in code it can't see. Same. Claude is doing its best, but you brought a screwdriver to drive in a nail.
Claude has no eyes on your screen, no live feed of the world, no memory of yesterday's conversation, and no access to anything you haven't pasted in. When you ask it about live facts (train times, prices, weather, current events) without giving it any source, it has to guess based on what it saw during training — and training ended a while back. The tool isn't broken. It just isn't the right tool for that job.
Pick the right tool for the shape of the question. Live facts → search engine. Look at a specific file → paste the file in, or use a tool that has access (Claude Code, Cursor). Specifics about your life → tell Claude what you know first, then ask. Real-time data → use a tool with web search turned on (Claude Pro, ChatGPT, or Gemini with search). And if a person can answer the question faster than you can prompt your way to it, ask the person.
A 30-second decision rubric
Before asking Claude, run this check: 1. Does the answer depend on what happened in the last 12 months? → If yes: you need web search, not Claude alone. 2. Does the answer require looking at a specific file, email, or page you have? → If yes: paste the content into the prompt, or use a tool with access. 3. Does the answer require knowing something about you Claude can't possibly know? → If yes: tell Claude first, then ask. 4. Could a person answer this in 30 seconds? → If yes: ask the person. If none of these apply, you're in Claude's territory. Proceed.
This is the most universal of the four failure modes. The fix transfers cleanly to ChatGPT and Gemini — the rubric is the same. The only difference is which tool you reach for in each case: Gemini's deep research mode is strong for long-form information gathering; ChatGPT's image generation and voice mode are strong where Claude is weaker; Claude is strong for writing, thinking-through, and code.
TAKEAWAY
The shape of skill
Notice what the four fixes have in common. None of them require new knowledge of Claude. They're all moves you make from your side of the conversation — what you ask, how you frame it, when you reset, when you pick a different tool. The skill isn't in knowing Claude better. It's in knowing yourself better when you're using it.
The yes-man dissolves the moment you stop asking for validation. The invention shrinks the moment you ask for a source. The circling stops the moment you start fresh. The wrong-tool problem disappears the moment you pause for thirty seconds before asking.
Get fluent in these four moves and you'll trust your tools properly — neither too much nor too little. That's the whole game.
Try it now
Three short prompts you can copy into Claude (or ChatGPT, or Gemini) right now. Each one practices one of the moves above. Notice how the answers change when you change how you ask.
Practice 1 — Pushback on something you wrote
I want you to push back hard on this. Don't be polite. If it's weak, tell me why. Three sharp criticisms beat ten compliments. [PASTE SOMETHING YOU'VE WRITTEN — AN EMAIL, A PARAGRAPH, AN IDEA]
Practice 2 — Confidence-flagging on a question of fact
Before you answer, flag your confidence. If you're inferring rather than sure, say so. If you can't point to a real source, say you don't know. Question: [SOMETHING SPECIFIC YOU'RE GENUINELY CURIOUS ABOUT]
Practice 3 — A clean reset on a conversation that went sideways
Fresh start. Forget what we were doing. Here's what I actually want: [ONE PARAGRAPH] Here are the constraints that matter: 1) [X], 2) [Y], 3) [Z] Here's where I am now: [PASTE THE LATEST VERSION ONLY] Help me push it forward.