I posted in our AI forum at work a while back about how I structure my projects so Claude Code agents can actually do useful cross-project work. One root folder with all my repos in it, around ten of them, a couple being monorepos. Inside that same root, a .claude folder with my agents. That way the agents weren't tied to one project. They could see across all of them.
For the kind of work I do, a lot of Elm to React migration, a lot of comparing how we did something in one place to how it should look in another, that wider view changes everything.
The post went over well. One colleague actually tried the setup and liked it, which made my day. Nobody else at the company is working this way yet, so seeing someone pick it up felt good.
Then later the same day I asked Claude if this was still the best way to do it.
It wasn't.
What the old setup got right, and where it fell short
The cross-project part was the right instinct. Giving the AI a view across all your work is miles better than pointing it at one project at a time, especially when you're doing anything that involves comparison or migration.
The problem was where I kept the agents. Putting them inside the root dev folder meant they only existed when I launched Claude from inside that folder. Open the terminal somewhere else and they were gone. I'd been working around that for long enough that I stopped noticing.
What changed
The user-level ~/.claude/ folder. Agents, rules, skills, memory, all of it lives there instead. It loads no matter where you launch Claude from or which project you're in. You don't have to "be in the right place" for your setup to exist.
That sounds small but it changes how it feels to work. The agents aren't a thing I have to remember to enter. They're just there.
The other piece I didn't know about before was the rules/ directory. Files in there auto-load in every session, without needing a slash command or a specific skill invocation. For things you always want applied, like coding conventions or style rules, that's the right place for them. Before, I had stuff tucked inside skills that only loaded when I asked for them by name. Which meant the setup worked, but only through one specific door.
And then a third version showed up
A few days later I came across a tip from Boris, the creator of Claude Code, about the --add-dir flag (or /add-dir inside a running session). It lets you add other folders to a session on demand, so you can pull in another repo without launching from a common parent. It also gives Claude permissions to work in that folder, not just read it.
Same problem my root-folder setup was solving, just a more surgical version. Instead of "see everything all the time," it's "see this specific extra thing for this specific session." For focused work where you don't want the noise of ten projects in scope, this is the better tool. For wider migration work where I actually want everything visible, the root folder still wins.
They don't replace each other. They're different shapes of the same idea.
The actual lesson
I shared something that was solid, and within a day I found a better version of it. A few days after that, a third. Not because I wasn't paying attention. Because these tools change fast enough that whatever clever setup you figured out a few months ago is worth re-checking.
I half-expected it when I asked. That's really the point. The habit of asking "is this still the best way?" is worth more than any specific setup, because the setup will keep changing. The habit doesn't.
I would say I ask this a few times a week now. Most of the time the answer is "yeah, still fine." But often enough there's a real adjustment, or a new piece I missed, or a pattern that has quietly become the better default. Enough that it earns the few seconds it costs.
A lot of people working with AI tools right now are in a loop where they figure something out, it works, and they stop checking in on it. That's how you end up confidently teaching people the 2025 way of doing something in 2026. You're not wrong because you're stupid. You're wrong because you stopped asking.
So this is the follow-up, not a replacement. The original post still stands. The setup I have now is better than the one I shared. And the next version, whenever I find it, will be better still.