Reference

When Claude
gets it wrong

It happens. Less often than people think, more often than the marketing suggests. The difference between a beginner and an intermediate is not that one gets wrong answers and the other does not. It is that one knows what to do next.

This page is a recovery guide. Not "how to prevent Claude getting it wrong" (that is most of the rest of the handbook), but "when it does, what now".

Noticing is step one

Before you can recover from a wrong answer, you have to notice it is wrong. The Intermediate guide covers the three reliable tells. As a refresher:

Add a fourth, which is easy to miss: the output matches the prompt exactly but does not match what you wanted. This is not Claude being wrong, it is the prompt being wrong, and you will see it in the output as "technically correct but useless".

The recovery escalation

Try these in order. Most wrong outputs are fixed by step 1. Most of the rest are fixed by step 2. By the time you reach step 5 you have a different problem on your hands.

Step 1: Add context

By far the most common fix. The output was wrong because Claude did not know enough. Figure out what was missing and add it. "Your previous answer assumed X, which is not true for us. Here is what is actually true: ...". Run it again.

This fixes the majority of wrong outputs. The prompt was thinner than the task required, and the model filled the gap with a guess.

Step 2: Ask what it needs

If you do not know what context was missing, ask Claude. "What would you need to know from me to do this properly?" or "What assumptions are you making that I should confirm?" Claude will usually come back with a list of the specific things it filled in. You supply those, run again, and the output is dramatically better.

This is underused. Most people, when they get a bad output, argue with it. Arguing almost never works. Asking what is missing almost always does.

Step 3: Start a fresh chat

If a conversation has been going for a while and the outputs are drifting, the problem may not be the latest prompt. It may be that earlier turns in the chat have anchored the model in a wrong direction, and it cannot easily unstick itself.

Start a new chat. Paste only the prompt and the context you actually need. Do not paste the previous bad answer asking for a rewrite. You will get a clean output that is not weighed down by everything you already tried.

Step 4: Try a different model

If the prompt is genuinely good, the context is complete, the chat is fresh, and the output is still almost right but not quite, upgrade the model. Sonnet is the default. Opus is the next step up. See the models page.

This fixes a specific class of failure: the task is at or just past the limit of what Sonnet can do. Opus brings more capacity to the same context and often produces a better answer without any prompt changes.

Step 5: Step away from Claude

If you have done all four of the above and the output is still wrong, the problem is no longer about Claude. Either the task is not a good fit for Claude (see When not to use Claude), or the task is good but the answer lives somewhere that needs a human doing it by hand, or the task is ambiguous enough that you need to talk to a real colleague before even writing the prompt. None of those are Claude problems. They are upstream of Claude.

Recognising this takes some practice. The temptation is to keep re-rolling prompts, hoping. Do not. If four honest attempts have not worked, the tool is not the right tool.

Things people try that almost never work

Arguing with the output

"That's wrong, do it again" rarely produces a better answer, because Claude does not know what you thought was wrong about it. Specific beats vague every time. "The third paragraph invented a number. Please rewrite without that sentence and flag anywhere you would have wanted to invent" is a useful rerun. "Try again" is not.

Adding "please be accurate"

Vague quality instructions do not do anything. Claude is already trying to be accurate. What works is specific constraints: "cite every number back to a source in the attached document", "if you are inferring something rather than reading it, flag it", "say 'not in the source' rather than filling in". Instructions beat suggestions.

Repeating the same prompt hoping for a different answer

Language models are slightly probabilistic, so technically you will get a slightly different answer the second time. It will not be dramatically different, and it will have the same underlying problem. If your first prompt produced a bad output, the fix is a different prompt, not the same prompt again.

Escalating tone

"I already told you this." "Stop doing X." "Why won't you listen?" These do not help. Claude is not being difficult on purpose, it is missing context. The energy you spend frustrated would be better spent figuring out what context is missing and adding it.

When a wrong output is actually the right output

Worth naming: sometimes the output is wrong and that is useful information. Not every bad output is a failure.

If you asked Claude for something and it produced something confidently incorrect, that is a signal that your prompt did not constrain the task well enough, and a real human taking on the same vague prompt would also have gone wrong, just more quietly. The bad output is diagnostic. Use it as a hint about what your prompt was missing, fix the prompt, and move on.

If you asked Claude for a creative rewrite and it produced something you do not love, that is not wrong, it is a preference mismatch. Ask for three versions with different angles. Pick one. This is not a failure mode, it is the normal shape of creative work.

If you asked Claude a factual question and it said "I do not know", that is correct behaviour, not a failure. The failure mode would be Claude guessing. Say "thank you for being honest" internally and move on.

A wrong answer is almost always a prompt problem. Fix the prompt, not the model.

For more: the cheat sheet has a short version of this page. When not to use Claude is the page to read before starting a task, not after it goes wrong.