How do you stop Claude making the same mistake twice?

You stop Claude making the same mistake twice by writing the fix down somewhere it will actually look. I do one of two things. If it's a task I'll repeat, I tell Claude to turn it into a skill and use that skill every time. If it's a final rule, I ask Claude to put it in the CLAUDE.md file, which it reads before every response. Mistakes that used to repeat for me now rarely do.

Near the end of a call, the person I was helping set up Claude told me something I hear a lot. The same mistake had happened twice. It used to happen to me too. Now it almost never does, and the fix is simple.

Which of the two you pick depends on whether you're dealing with a task or a rule.

What do you do when it's a task you'll do again?

Any time you do something repetitively, or you get a really beautiful output, that's your signal. Don't let that work disappear when the chat ends.

Say you've just had a long chat with Claude working through something properly. You got it to do the job the way you want. Just say it: "Hey Claude, we will probably do this again. Turn this into a skill, and use that skill whenever I ask you to do this."

It will do that. It will tell you it has saved the skill. Next time, you don't explain everything from scratch. It already knows how you like it done.

What about rules it should never break?

The second thing is for rules. Final rules. The things it should always do or never do, no matter what.

Here's one of mine. Never message anybody, even if I tell you to send. Send means draft. Never send. I could tell Claude to open WhatsApp Web and send a reply for me. I have told it not to, and I put that in as a rule. That mistake happened to me before. Now it never does.

Why? Because whenever something is a final rule, I say, "Hey Claude, please put this in the CLAUDE.md file."

The CLAUDE.md file is a set of rules your Claude reads before any response, ever. So the rule isn't something it has to remember from an old chat. It's in front of it every time before it answers you.

I'm at 145 rules now. That's because every day I've added some rule or the other. Things like: make sure you look at this before you answer me. Or: never use an em dash in any output, ever. Even if you're only talking to me.

And it holds. I don't see em dashes in the emails or messages it writes for me anymore.

When you add a rule, it tells you. It will say something like "I added this as rule number so and so." So you know it's there for next time.

How do you catch the mistakes in the first place?

You read what it did. That's the whole habit.

When someone sends me a request, I'll often take a screenshot, drop it into Claude and ask it to make the change and tell me what to reply. Before anything goes out, I read what it did. If it's good, I copy the reply and I'm done. If there's an error, I tell it straight: you made a mistake, do it again. Then I tell it what to message.

That checking is what feeds the skills and the rules. Every day, go and read what Claude did and check if it did it right. Each error you find is something you can fix for good.

And it adds up. After making, say, 100 errors on reading one kind of document, Claude will be a master at reading that document. It will not make those mistakes again. Keep finding errors for long enough and the people checking its work end up with very little to check.

Can Claude find its own repeated mistakes?

Yes. I voice type everything through Whisper, so all my voice typing stays in one place.

Every week, Claude goes through my Whisper transcripts and looks for the mistakes I have kept telling it to stop making. Then it adds them to my skill. So every week it gets better, without me sitting down to list my complaints.

You can set this up the same way. Tell Claude you want it to become better at serving you. Explain your workflow: this is how I work, I voice type everything, everything is in Whisper, you can go and read it. Then let it do the review.

Why does this habit matter?

So the habit is small. When you finish something you'll repeat, make it a skill. When you catch a mistake you never want to see again, make it a rule in CLAUDE.md. Read what it does every day, and let it review its own repeated mistakes every week. Do that and your setup gets better every day, without you repeating yourself.

“The CLAUDE.md file is a set of rules which your Claude reads before any response ever.”

Said on the call, 2 October 2026