You put your config in my config
not all sharing is caring
I’ve been thinking a lot lately about collaboration, and how important it is to me. In my view, pretty much any project worth working on involves working together — no-one is an island, after all. There’s so much good that comes from working together. Sure, splitting up the work is nice, but more than that I value the influences we can have on each other, and especially the celebrations we can have because we’re not all the same. We have our own ways of seeing problems, coming up with solutions, and working through them.
But there are two sides to this. Once, in my consulting career long long ago, we worked with a client that had one guy who used Dvorak for his keyboard layout. And he said he’d pair with anyone as long as they changed their keyboard layout to Dvorak. That’s a hostile and bad-faith sort of offer, because he really didn’t want to work with anyone else. (And really, it shouldn’t even have been an issue. You can make quick-switch shortcuts if it’s important to you.)
Like anything else, I want to find a good middle ground between “everyone go hog wild and do whatever you like” and “everyone in complete lock-step, and boy oh boy I really wish you were just replaceable cogs in my machine”. I’ve had plenty of experience with this over the years, but what I’m running into lately involves (of course) LLMs, and it feels (of course) like a big shift — something new, different, and concerning.
I am continuing my journey with Claude, and I’ve gotten somewhere with my personal setup that works for me — and continues to evolve, as I learn and grow(?). I think the way Claude handles layering with its settings and instructions is very useful, letting you have a baseline in your personal ~/.claude/, project-specific things in a project’s .claude/, and even having CLAUDE.md files anywhere in the folder hierarchy.
That’s great. But this is also where I’m running into trouble.
We’ve done plenty of talking internally about the different Claudes you may get when you’re trying to work, how you start with “dum-dum Claude” when you have no setup whatsoever, and then you set up more and get used to working in a certain way, with that assistant you “know”. But you have to be careful, because working in a particular project can turn your assistant into something totally different. In just the past few months, I’ve had:
- skills changed from under me — the same ones, the same names, but older versions that don’t handle the various corner and edge cases I’ve found
- MCPs activated on me, asking me for setup I don’t need
- hooks installed for workflows that thankfully I don’t use, but now feel I have to avoid entirely
- permissions for tools changed on me — either denying things I use regularly or allowing things I expect to be asked about
All of these are examples of personal preferences being set as shared configuration.
It can be hard to see the dividing line, and there’s probably more grey area than is really comfortable. Linting can feel like someone’s taste, but it gives everyone one consistent style to read. A bin/setup script can feel presumptuous when it uses brew install, but it gets you up and running quickly. Git hooks can keep you from publishing (or even committing) secrets, but impose some heaviness and slowness on your workflow. LLM skills can automate complex workflows (and yes, skills did show up on my list of grievances above). All of these things keep everyone on the same page, and I certainly value consistency a great deal.
But there is a division between personal preferences and shared configuration. Shared configuration is fact, rather than preference. Facts get committed and shared, preferences don’t.
Some examples:
| Fact | Preference |
|---|---|
| how to build, test, lint | editor settings, review habits |
| dependencies (ruby version, gems, system needs) | use chruby and Homebrew |
| conventions, domain terms, gotchas | plan extensively, work in parallel with worktrees, never stash |
what an assistant must never do (read .env, deploy to prod) |
what’s fine to do without asking |
So where do these things go? Your ~/.claude has the behavior that applies to all of your projects, while a project’s local settings have what you want for this project only. The project’s env file has what makes sense on your machine (like which port to use rather than fighting over 3000), while the committed example file tells you what needs to be set. Your global gitignore keeps everyone else from dealing with your editor’s files in any way.
None of this means you should keep your preferences to yourself. I publish my various setups, but they’re meant as somewhere between reference and inspiration, not as something for others to install. If you think one of your preferences is worth sharing, propose it. The team may agree, which makes it a convention, and that’s a fact. The trouble starts when one person’s preference ends up being shared such that everyone else has to opt out.
And I’ll admit that I’m far from perfect here. I’ve made large transgressions like unthinkingly adding things to a project that push it towards “works on my machine”, and I’ve made small transgressions like adding configuration specific to my editor, or even trying to be nice and adding gitignore entries covering that configuration (rather than putting those in my personal gitignore).
I don’t think anyone committing their workflow is trying to be that guy. They’re trying to help, just as I was with those gitignore entries. But it ends up feeling the same, like someone’s changed my keyboard layout to Dvorak, and I didn’t even notice until C oyapy.e yflcbiv
If you’re looking for a team to help you discover the right thing to build and help you build it, get in touch.
Published on November 13, 2026
