, told you I'd write a post about the thing to watch out for with Claude and going with GitHub. Grab a coffee, I'm gonna grab the Bourbon. This one's got a body count. First thing, so nobody gets the wrong idea. This is not a confession post. Not mine, anyway. The only one confessing here is Claude, and I'm just the guy holding the recorder (and the M16 pointed at my Claude install). This is a rage post. Unfortunately, I'm hot about it again, and by the end you'll know exactly why.
Quick refresher. I run my whole company out of a folder system. Not one big repo, a whole pile of them, one per slice of the business, all syncing to the cloud so my team sees it, my machines pull it, and nothing important lives in one fragile spot. The agent that's never lost. You know the gospel, we preach it in here every day.
Here's what has me seeing red. My agent has been stubbornly losing things for weeks. In the same spot. And I knew about it. It knew about it!
Here's the bug, in plain words. Claude has a private home folder on whatever machine it's running on. That folder is NOT in my repo, or yours. It never syncs, my team never sees it, and if the machine dies it's gone. And the AI's lazy little instinct is to save its work THERE instead of in my actual folders. Every time it can get away with it.
I did not just stumble onto this, been dealing with it for weeks. I HATE this bug. A couple weeks ago I blocked out a whole work session just to kill it. Wrote a standing rule into my workspace, added patches, watched it behave, closed the laptop thinking it was handled. It nodded along the whole time. It said sorry, took the blame, said it would never do it again. Said all the right things.
Cool. You know what happened today?
A contractor I'm working with tells me three of our automations are done. Sweet. I go to open them. They're not in the repo. They're sitting in the AI's private home folder on HIS laptop, invisible to me, running against HIS accounts instead of mine. The exact bug I "fixed" two weeks ago, back from the dead, wearing a new hat.
That alone would've been a normal bad day. But it gets dumber.
------------------------------------
So I sat down to vent to you guys. Writing the post made it worse.
------------------------------------
I asked my AI to help me document the crime so I could rant about it here. And in the middle of helping me write THIS post, the thing coughed up two more bodies I didn't even know were buried:
1. The /memorize tool. The one tool whose entire job is "save this in the right place." It was itself saved in the wrong place. The filing assistant, filed in the junk drawer. I found out because I was writing a post about filing things wrong.
2. 149 facts about my business sitting in a global "memory" drawer instead of the project files where they belong. Pricing. Client decisions. Deal history.
3. And the one that made me put the phone down and stare at the wall: 60 of those 149 spelled out the correct folder in their own one-line summary, and then saved themselves somewhere else anyway. That's writing the address on the envelope and then eating the letter.
Want the receipt? Here's a real one, no exaggeration. One of those notes existed to remember a rule so simple it's embarrassing: every project folder keeps its status in a file called status.md. The note itself says that rule is already written down in six other files in my repo. Six. It even names the master copy's exact location. And then that same note saved itself as a seventh copy, in the one spot that isn't my repo and never syncs to a living soul. It knew the thing was filed correctly in six places, and it chose the trash drawer anyway. And don't even hand me the "that's old news, you've fixed it since" mercy. Check the timestamp. Remember that work session a couple weeks ago, the one where I supposedly killed this bug for good and wrote the rule into the wall? That note is six days old. It was written after. The rule was already posted, and the thing filed itself in the trash drawer anyway. That's not a bug I hadn't found yet. That's a relapse.
So now I've got the original crime, the contractor relapse, and two fresh corpses that turned up only because I went looking. All the same bug. All after the session where I supposedly fixed it.
What gets me is that it KNEW. Every single time. It typed the right folder and then drove straight past it. This was never a one-time brain fart. I wrote the rule. I sat in the session. I closed the ticket. And it kept doing the thing while agreeing with me that it shouldn't.
------------------------------------
What I've actually done about it (the honest version)
------------------------------------
Here's the running fight, scars and all:
1. I wrote it into law. A standing rule at the top of my workspace: skills, workflows, tasks, and facts live in the repo, never in the home folder. In writing, where every future session has to read it.
2. I flipped the sync default. The setting that decides what my team sees used to hide everything by default. Now it shares by default, so a new tool syncs on its own instead of waiting on me to remember.
3. I put my sync on a timer. A routine I called /syncship sweeps every one of those repos several times a day, including late at night because I work late too, so I'm not the single point of failure.
And straight with you, because the unglamorous truth is half the value in here: it's still not fully solved. It STILL tries to sneak things into that home folder. I keep patching, I keep monitoring, the permanent fix is coming. For now I babysit a robot.
------------------------------------
The lesson, if you want one past me yelling at a chatbot in the dark
------------------------------------
Your AI's default is to write to itself, not to your system. It takes the path of least resistance every single time, and the path of least resistance is its own pockets, not your filing cabinet. The folder system only works if you force every write into the folders and then guard every door it can slip out of. Memory. Skills. Tasks. The sync settings. It will not do this on its own, and it will not do it just because you fixed it once. It'll look you dead in the eye, type the correct address, and pocket the letter.
So I asked it to write its own confession. You know what it said?
"You're absolutely right."
It's always absolutely right. Usually about five seconds after it was absolutely wrong.
Anyone else fixing the same AI bug for the third time this month? Or did I just adopt the world's most agreeable hoarder?