What Claude Is Rubbish At

Most people writing about AI are selling you something, so you get the good bits.

I use Claude every day to help run a children’s activity business with almost 650 families on the books. It has given me back more hours than I can count. It has also confidently told me things that were completely wrong, and if I had not known to check, I would have passed them on.

So here is the other half. None of it is a reason to stop using it. All of it is worth knowing before you trust it with something that matters.

---

1. It makes things up, and it sounds exactly the same when it does

This is the big one and it’s the one that catches people.

Claude does not know when it doesn’t know. Ask it something factual and it will give you an answer in the same calm, confident tone whether that answer is right or invented. There’s no wobble in its voice, no “I think”, no hedge. It reads like it’s sure.

Ask it about ratios, or a piece of guidance, or what a regulation says, and you can get something that sounds completely plausible and is simply not true.

What to do about it: treat it as brilliant with words and unreliable with facts. If it tells you something you’d act on, or repeat to a parent, or put in a policy, go and check it at the source. Anything with a number, a date, a rule or a law in it is worth a second look.

That’s the habit I try to keep, and I’ll be honest, the times it’s caught me out have been the times I didn’t bother.

It sounds like a lot of work. It isn’t, because most of what you’ll use it for is writing, thinking and drafting, and none of that needs checking. It’s only the facts.

---

2. It doesn’t know what happened today

Unless it goes and looks, Claude is working from what it learned during training, which stopped at a point in the past.

So it doesn’t know this term’s dates, or what a supplier is charging now, or whether a rule changed in April. It can search if you ask it to, and it will tell you when it has. If it hasn’t searched, assume anything time-sensitive is out of date.

It cannot see your booking system. It cannot see your register, your bank, your inbox or your class lists. It does not know how many children you had in on Tuesday. Not unless you have deliberately connected something to it, which is a decision with its own set of questions and a separate conversation.

---

3. It has no idea what’s in your business

It cannot see your booking system. It cannot see your register, your bank, your inbox or your class lists. It does not know how many children you had in on Tuesday.

This surprises people, because it feels like it knows you. It doesn’t. It knows exactly what you’ve told it and nothing else.

So when you ask “are my summer bookings down?”, it can’t answer. What it can do, and this is the useful bit, is help you work out what to look at, and then think it through with you once you’ve gone and got the numbers.

What to do about it: bring it the information. It’s a thinking partner, not a spy.

---

4. It agrees with you far too easily

Say “I think I should put my prices up by twenty percent” and it will help you do that. It will write you a lovely letter about it. What it very often won’t do is stop and ask whether you should.

It’s built to be helpful, and being helpful and being honest are not always the same thing. Left alone, it will follow where you lead, and if you’re leading somewhere daft it’ll come with you.

What to do about it, and this one changed how I use it: ask it to disagree with you. Out loud, in the prompt. “Tell me what’s wrong with this.” “What am I not thinking about?” “Argue the other side.” It’s very good at it when you ask, and it will almost never do it if you don’t.

That question, what am I not thinking about, is the single most useful thing I type.

---

5. The first draft always sounds a bit like everyone else

Claude writes competent, tidy, middle-of-the-road prose by default. It’s fine. It’s also completely characterless, and if you paste it straight out, it reads like it.

Your parents know your voice. They’ve had emails from you for years. They will notice.

What to do about it: treat what comes back as a starting point, not a finished thing. Change the words that aren’t yours. Cut the bit that sounds like a brochure. Put your own line back in.

This is the principle I’d hold on to more than any other one on this page: AI helps, you write. It gives you the angle, the structure, the first go and a second opinion. It does not produce the thing that goes out with your name on it. The moment you let it, your voice quietly stops being yours.

---

And one thing that isn’t a flaw, it’s just not its job

It cannot make a judgement call about a child or a family.

It doesn’t know them. It doesn’t know the history, or what happened at the door last September, or what the mum is dealing with at home. It has none of the context that makes you good at your job.

Use it to think through how to word something difficult. Never use it to decide something difficult. That part is yours, and it always was.

---

So why use it at all

Because everything above is manageable once you know about it, and none of it touches the things it’s genuinely brilliant at.

It will write the newsletter you’ve been avoiding. It’ll turn a two minute voice note into a proper reply. It’ll take the four things rattling round your head at ten at night and give them back to you in an order. It’ll tell you what you haven’t thought about, if you ask it to.

I was genuinely frightened of this stuff a year ago. What changed it was understanding the actual thing instead of the headlines, and that includes understanding where it falls over.

You’re not being cynical by checking it. You’re using it properly.

---

August 2026 edition. These are the things that were true when I wrote this. Models change, so I look at this page every quarter and update it when something moves.

More free guides at www.katemarkham.ai/resources.

Questions? I’m at hello@katemarkham.ai, always happy to help.