Claude Code for Knowledge Work

By Peter Kaminski

This is an early field note from onboarding non-software developers onto Claude Code to work on large document sets rather than code. It’s early, but already useful: a concrete pattern is emerging, along with some sharp edges worth naming.

The goal is to make sustained progress on complex written artifacts — building out the content for a hundred-page website, things of similar scope.

I’m sure later this year we’ll have something as powerful working on your computer as Claude Code, but without the prickly and sharp edges of Claude Code. We’re not there yet, so I don’t think Claude Code is for everybody right now. (Claude Cowork will possibly be that successor, but it’s not quite there yet.)

And yet, if someone is an Obsidian user, understands “files,” and has some forbearance for tech, I think they want access to the stratospheric productivity and skill enhancements that powerful agentic AI like Claude Code gives you now, rather than in 6-12 months. The UX is tolerable, especially with a bit of handholding to get started.

The pattern I’m liking a lot right now:

  • Running Claude Code in VS Code, rather than in the terminal or on the web/desktop. The terminal is something only geeks can love. The web/desktop versions are okay to use, but it’s a pain to get the git branch from its virtual computer back to your own computer, unless you have a pretty good model in your head of git-based collaboration.
  • Saying that VS Code is a lot like Obsidian, just less friendly. :-) The folks I’ve onboarded are familiar with Markdown and Obsidian, and Obsidian is just a Markdown IDE, so this goes over pretty well. I’m even saying to have both VS Code and Obsidian up at the same time, pointed to the same folder / vault / repo. Use Claude Code in VS Code, use Obsidian to navigate, view, and edit, because it’s more friendly. (I also use Typora in the mix, but adding yet another tool just to optimize UX and user delight can be overkill here.) SOMEDAY, Anthropic or someone (you? me?) will build Claude Code into Obsidian, but it’s not too bad having both up for now.
  • Git via English-language commands. I think version management is one of the hidden wonders of the software development industry, and I love it when non-software developers can use it. It adds a ton of safety and optionality to the development process of, say, a hundred-page website, or that novel you’re working on, or a big set of legal agreements. And with Claude Code, you don’t remember git commands, you just tell Claude Code to make snapshots (commits) and to back up to the cloud (push). We’ll get to git-based collaboration through Claude Code later, but version management and agentic AI is a match made in heaven.
  • Billing via Claude Pro (or Max) account rather than API. It’s a great flat-rate deal, and your friend will probably figure out how to use enough Claude Code to make US$20/month worth it.

The biggest hurdles?

  • Authenticating your computer to GitHub. This isn’t impossible or anything, but it’s fiddly.
  • What to do about the safety pop-ups. So far the folks I’ve been working with have enough computer literacy that they sort of know the possible problems, and at least have backups and know not to download weird stuff as much as possible. A hard sandbox would help, but then you don’t have your files on your computer, in the sandbox. It’s a power/safety/convenience trade-off thing. I’m looking forward to an awesome solution to this, but in the meantime, some of us are going for it, the same as software devs have been doing in 2025.

More to come! — Pete