A bad output is a draft, not a verdict.
Claude is going to get something wrong. The first time it happens, most people quietly close the tab and decide AI isn't for them.
That's the wrong move. Bad outputs aren't proof Claude is broken. They're a signal that the prompt, the context, or the ask needs adjustment. This guide is how to read the signal and fix it.
Spotting which kind of wrong it is.
Most bad outputs fall into three categories. Knowing which kind you're looking at tells you what to do next.
Hallucination
Claude states something confidently and it isn't true. A made-up source. A statistic that doesn't exist. A quote attributed to someone who never said it. The tell: Claude sounds the most confident exactly when it's wrong.
Refusal
Claude declines to help with something that should be helpable. The tell: a response that explains why it can't do the thing instead of doing it. Often longer than the answer would have been.
Weird tone
The answer is technically right but sounds off. Too formal. Too casual. Too long. Too eager. The tell: you read the output and your gut says "I wouldn't send this."
All three are fixable
You don't need to start over. You need a follow-up. The next chapter is the three follow-ups that handle most cases.
What to type next.
When you get a bad output, you almost never need to start over. You need to send a follow-up. These three handle most cases.
That sounds confident, but I want to verify. For each specific claim, fact, or source you used in the answer above, tell me how confident you are (high, medium, low) and where you got it. If you can't source something, tell me that too.
Paste this any time Claude gives you a fact or source you'd want to check.
Claude will almost always flag the parts it made up. The trick is asking. Until you ask, it doesn't volunteer.
You declined that, but I think there's a way to help here that I haven't explained well. Here's what I'm actually trying to do: [the real reason in plain English]. Given that, can you take another look? If there's still part you can't help with, tell me what part specifically.
Paste this any time Claude declines something you genuinely need help with.
The key is the real reason in plain English. Not the request again. The why behind it. Most refusals come from misread intent. Naming the intent fixes it.
The content is right but the tone is off. It reads like [too formal / too eager / too generic / too long]. Rewrite it in the same length but with this tone: [your specific fix]. Keep all the actual information, change how it sounds.
Paste this any time the answer is correct but you wouldn't send it.
The specific fix is what makes this work. "Better" is too vague. "More like a colleague catching up over coffee, less like a press release" is what Claude can actually act on.
Three signs the chat is cooked.
Follow-ups fix most things. Sometimes they don't. These are the three signs the chat is cooked and you should start a fresh one.
First , when Claude has given you the same wrong answer three times in a row even with corrections. The chat has anchored on the wrong frame. Open a new chat and rephrase.
Second , when the conversation has drifted so far from your original task that even you've forgotten what you came in for. Open a new chat with just the cleaned-up ask.
Third , when Claude is being weirdly stubborn about something. Sometimes the way the chat got framed early makes it hard to recover. New chat, new framing.
The setup that catches problems early.
Most bad outputs come from one of two upstream problems. The prompt didn't say what "good" looks like, or it didn't include the source material Claude needed to be accurate.
For accuracy: paste the source. Don't ask Claude to remember a study. Paste the study. Don't ask Claude to summarize an article from memory. Paste the article.
For tone: tell Claude what bad would look like, not just what good would look like. "Don't sound like a chatbot" sets a clearer floor than "sound like me."
Common traps.
- Trusting Claude on specific facts, dates, numbers, or sources without asking it to flag confidence. Always ask.
- Accepting a refusal at face value. Half of refusals dissolve once you explain the why.
- Editing the output by hand when a follow-up would do it faster. The whole point of Claude is the follow-up.
- Starting a new chat too quickly. Try one good follow-up first.
- Starting a new chat too late. If the same bad answer comes back three times, leave the chat.