Journal · The AI Architect · 2026-08-13
One job, one folder. Give your AI the drawer the task needs, not the keys to the whole cabinet.
One job, one folder. Give your AI the drawer the task needs, not the keys to the whole cabinet.
The first time I gave an AI assistant access to my entire file system, I told myself it was about efficiency. No more copying files into folders, no more re-uploading documents. Just let it see everything and figure out what it needs. Three weeks later, it referenced a client’s tax documents while helping me draft a blog post. Nothing catastrophic happened. But I sat there thinking about all the ways it could have gone wrong, and about how casually I’d handed over that risk for the sake of saving myself a few clicks.
That’s the real tension. It’s not that broad access is obviously dangerous and narrow access is obviously safe. It’s that broad access feels convenient right up until the moment it isn’t, and by then the habit is already formed. Most people don’t sit down and decide to give their AI tools sweeping permissions. It happens gradually, one “just this once” at a time, until the AI has a master key and you can’t quite remember handing it over.
Convenience now, cost later
The pitch for broad access always sounds reasonable in the moment. It’s faster to grant an assistant access to your whole drive than to curate a folder for every task. It’s easier to connect an email plugin to everything than to scope it down. But convenience purchased today is often paid for later, usually at a worse exchange rate. A tool that can see everything will eventually be asked to act on things you didn’t intend it to touch, and “the AI made a mistake” is a much harder sentence to say calmly when the mistake involves your entire client list instead of one project.
Try this: before your next AI task, spend thirty seconds asking what the smallest possible set of files or data this actually requires. Then create a folder with just those things, and point the tool there instead of at your broader drive. If the task is drafting a report, it needs the source documents for that report, not your entire archive of past projects. Treat that folder as the whole world for the duration of the task. If the AI can’t do the job with what’s in the drawer, that’s useful information too, not a reason to hand over the cabinet.
The drawer as a design choice, not a limitation
There’s a temptation to think of scoped access as a compromise, something you accept because full access would be risky. But the drawer isn’t just a safety measure. It changes what the AI actually does. When a model can see one relevant folder instead of ten thousand files, its answers get sharper, because it isn’t guessing about what matters. Ambiguity is the enemy of good output, and unlimited access is a firehose of ambiguity. A tight, well-labeled drawer is a kind of instruction in itself. It tells the AI, implicitly, “this is what matters here,” which is a favor you’re doing for the model as much as for yourself.
This also makes mistakes easier to trace. If something goes wrong and the AI only had access to one folder, you know exactly where to look. If it had access to everything, you’re stuck reconstructing what it might have seen, which is a much worse position to be in when you’re trying to fix a problem quickly.
Building the habit, not just the rule
Knowing the principle doesn’t mean you’ll follow it under deadline pressure. The habit has to be built the same way any habit is: with friction removed from the right choice and added to the wrong one. Set up your workspace so that creating a task-specific folder is the path of least resistance, not an extra chore. Keep a template structure ready. Get in the practice of asking, at the start of any new AI task, “what’s the drawer here?” before you ask “what’s the prompt?”
Try this: for one week, before starting any AI-assisted task, write down in one line what folder or data source you’re about to grant access to and why that’s enough. It’s a small piece of friction, but it turns an invisible habit into a visible decision, and visible decisions are much easier to correct when they’re wrong.
None of this requires distrust of AI tools themselves. It requires distrust of your own convenience-seeking, which is a much more reliable villain in most of these stories. The drawer isn’t a limitation on what your AI can do. It’s a limitation on how much can go wrong when it does something you didn’t expect.
This is just one piece of a larger way of thinking about scope and control that I get into more fully in the book, particularly how it plays out across different kinds of tools and tasks. But the folder is a good place to start, because it’s a habit you can build this week, on the very next thing you ask an AI to help with.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.