Well I just had to come on here and basically have a rant! I’m wondering is this just me this happens to?
I must have told Claude the same thing at least seventeen times.
Not seventeen times in one conversation. Seventeen times across multiple sessions, on different days, in what I can only describe as the world’s most patient argument with a very confident machine that was absolutely, categorically, provably wrong.
The line in question was this: “contAIn is a project management methodology for anyone building anything with AI.”
And every time it appeared, I corrected it. I’d say. “Here’s the correct version. Please use this one.”
“Of course,” Claude would say, with the digital equivalent of a polite nod. “Noted. My apologies.”
Next session, and there it was again. Same wrong line, same confident delivery, same amnesia, no recollection we’d had this conversation before.
Seventeen times.
Now, you might be thinking: Sam, love, that sounds like a Claude problem, and you’d be partially right, I didn’t ASK it to tell me what was wrong, a bit like that mate who just says fine when you know it is far from fine.
It is entirely, completely, one hundred percent a me problem. And the reason I know that is because when we finally sat down and pulled the thread, properly, all the way back to the source, we found the line wasn’t just in Claude’s session output.
It was in three places simultaneously.
It was in my memory settings, sitting there quietly feeding the wrong definition into every conversation before I’d typed a single word. It was in my Project Knowledge, a document I’d uploaded months ago that Claude loads at the start of every session like a very diligent student who has done all the pre-reading. And it was in the Skills file itself, the one which was doing exactly what it was designed to do, reliably, consistently, automatically producing the wrong line every time.
I had corrected the output seventeen times without ever touching the source.
Think about that for a second.
Every correction I made vanished the moment the session closed. The next Claude that opened that conversation had no idea the correction had happened. It just read the sources, found the wrong line in three separate places, and produced it again with complete confidence.
Here’s the thing about AI that the “60 secret commands that will transform your business” crowd never mentions: it is only as good as what you’ve built around it.
Claude didn’t invent that wrong line. I gave it to Claude, put it in the memory settings, uploaded the document that contained it, packaged it into the Skills file, and then spent seventeen sessions being baffled as to why Claude kept saying it.
The machine was doing exactly what I told it to do, the problem was I’d told it the wrong thing, in three different places, and then kept correcting the symptom instead of the cause.
This is what I mean when I say most people are only running 20% of their AI setup. Not because they don’t know enough prompts, haven’t found the right shortcuts, or are missing some hidden feature that would unlock the other 80%. But because nobody has told them that what you feed the machine matters more than how cleverly you ask it questions.
Your AI has sources, it reads them every session. If the sources are wrong, the output is wrong, and no amount of in-session correction will fix it, because the next session starts from the sources not from the conversation you just had.
This is not a Claude problem. This is not an AI problem. This is a you-haven’t-set-up-your-system problem, and it has a very boring, very unsexy, but very effective solution.
We now have a rule out of that session, it’s called the three-point source update. Any time a correction is made to a locked line, a definition, anything that needs to stay consistent, it gets fixed in three places before the session closes: memory, Project Knowledge, Skills. All three, there and then because it’s critical consistency.
The machine didn’t need to be smarter, I needed to be more structured. The whole contAIn methodology exists because of exactly this kind of moment, not the dramatic AI failure, nor the headline-grabbing hallucination, but the quiet, grinding, seventeen-sessions-long drip of wrong output that kept happening.
You don’t fix AI by arguing with it. You fix AI by fixing what you fed it in the first place, and if you’re wondering whether this applies to you: when did you last check what your AI actually thinks you do for a living?
Go ahead ask it, you may be surprised by the answer.
configure YOUR system. contAIn™ the chaos. control YOUR outcome.
This article was originally published on Substack.