When not to
use Claude
Most of the handbook is about what Claude is good at. This page is the honest counterpart: tasks where Claude is not the right tool, and what you should reach for instead.
Every handbook like this one has a version of "here's what it does brilliantly". Almost none of them say where it falls over. The ones that do not tell you the limits are setting you up to burn a week on a task you should never have started with Claude in the first place.
This is my list. Some of these have workarounds (an API call, a tool, a different model). For the chat window most readers will use, they are situations where the right answer is "do not use Claude".
Maths with precise answers
Language models are famously bad at arithmetic. Claude has got a lot better but it is still not the thing you want calculating your VAT return, summing a column of numbers, or working out compound interest. It can look right and be off by the last decimal in a way you will not spot.
A calculator, a spreadsheet, or ask Claude to write a tiny script that does the calculation and shows you the working. The second option is genuinely great. "Write me a Python one-liner that computes X from these inputs" is a perfect Claude task. Running the one-liner is not.
Real-time or current information
In a plain chat without web search enabled, Claude does not know today's date, today's news, today's share price, or what was posted on a competitor's blog this morning. Its knowledge has a training cut-off and that cut-off is not today. If you ask "what did the Bank of England do yesterday", Claude will either admit it does not know or, worse, confidently tell you something that was true six months ago.
A news site, a search engine, or Claude with web search or a connector turned on. The chat Claude knows a lot, but it does not know what is happening right now.
Tasks that need a single precise known answer
"What is the minimum wage in the UK right now", "what is the exact wording of section 4.2 of our MSA", "what is the capital of the country with the second largest GDP in 2024". These are tasks with one right answer that lives somewhere authoritative. Claude might get them right. It might get them subtly wrong. You will not know without checking, and the checking takes longer than skipping Claude and going to the source.
The authoritative source. Government site, the actual contract, the company's latest filings. Save Claude for the kind of task where judgement, drafting, or synthesis is the hard bit, not where the hard bit is remembering.
Anything load-bearing without your eyes on the output
This is not a hard "never" but a hard "not without checking". Claude can draft an email you are going to send to a client. Claude can draft a contract clause you will negotiate with. Claude can produce numbers that will end up in a board pack. In every one of those cases, the final responsibility is yours, and the review you do on the output is the bit that protects you. Do not ever send, publish, commit, or quote something Claude produced without reading it carefully. It is probably fine. You still read it.
Claude, but with the specific discipline of reading every word before it leaves your machine.
Personal data belonging to named individuals
Performance reviews tied to specific people. Medical details. Financial details tied to a named customer. Anything where the wrong person seeing the output would be a real problem. The team account is covered by proper enterprise terms, so this is not a "Claude will leak it" concern, it is a "use judgement" concern. The rule: do not put anything into Claude you would not put into an email to a non-privileged colleague. If the answer for a specific piece of content is "I would never email this", do not prompt with it either.
The normal non-AI tools you would have used before Claude existed, or Claude with names redacted and specifics abstracted. If unsure, ask your team lead.
Deciding anything
Claude can lay out options, synthesise inputs, and suggest a direction. It should not be the thing that makes a decision with real consequences. If you find yourself typing "should we hire this person" or "should we take this deal", Claude can help you think but it cannot take responsibility, and you should not pretend it has. The synthesis is useful. The decision is yours.
Claude as a thinking partner, your own judgement as the decider. The line between those two is the thing you are paid for.
Tasks that need perfect consistency at scale
If you are running the same task a thousand times and every output must be identical in format and correct in every particular, a plain chat with Claude is not it. Language models are probabilistic by design. Two runs of the same prompt on the same data can produce different outputs. For most work this is fine, even desirable. For "I need exactly this structure 10,000 times and one error per thousand is a problem", you want a proper pipeline, not a chat window.
The API with structured outputs, a proper validation layer, and evals. This is the Advanced tier. The Advanced guide covers it. Or, if you are not building a pipeline, accept that a chat is the wrong tool and use a normal script instead.
When you are looking for emotional support as a shortcut
Odd one to end on, but worth naming. Claude is good at sounding warm, and it is tempting to vent to it when you are stressed about work. That is fine, occasionally. What it is not is a substitute for a real colleague, a real friend, or a real therapist. It does not know you, it has no history with you, and it will tell you what you want to hear if you give it too much rope. Use it for the drafts, not for the relationships.
People. The hard version of this advice is: the person you are avoiding talking to is usually the one you most need to talk to, and a chatbot is not the workaround.
The short version
Claude is best at tasks where:
- Judgement and drafting are the hard part, not remembering.
- There is no single correct answer, just a better and a worse one.
- You can read the output before it matters.
- You are bringing the context and Claude is bringing the craft.
It is worst at tasks where:
- Arithmetic precision matters to the last decimal.
- The answer depends on information Claude does not have.
- Nobody will read the output before it goes somewhere important.
- You are hoping it will take a decision off your hands.
Knowing when not to use a tool is most of knowing how to use it.
Next: When Claude gets it wrong is the other honest page. When you do use Claude and the output is not what you hoped, here is what to do next.