Claude Code + Obsidian
Dear Scholar,
If you have opened social media in the last couple of weeks, you probably could not escape the countless "Claude + Obsidian" posts, promising an "AI second brain".
It started with a post by Andrej Karpathy, one of the founders of OpenAI (ChatGPT), who described a system for using AI with local files and Obsidian. From there, people quickly created all kinds of systems to manage local files with AI.
The idea: store everything you know, read and think about locally (using Obsidian), then use a large language model (Claude Code or Cowork) to process and synthesise these files. (Don't know what Obsidian is, start here).
The promise: Instead of uploading a few files to AI, you can now work with an unlimited number of files. Imagine being able to synthesise ideas from 100s of PDFs and your notes, from years of research. The idea is called "File over app".
The problem: You can't just tell AI to read all your files. Giving more context to an AI model leads to so-called "context rot", making replies worse. Just like a human who gets overwhelmed with too much content.
The solution: ...is somewhat complicated; you have to create documentation for your AI and have it operate in small, concise steps that do as little as necessary to maintain a clean context while still getting the job done.
That is where Claude Code comes in. Claude Code is such a system which can work with large "piles" of content. It is built on the premise of an AI agent (i.e., an AI that solves a problem in multiple steps rather than answers a question), has access to your local files, and offers many ways to control what happens inside this machine. It is a remarkable system, but it takes a little time to learn.
The real problem: Last week, Debby posted in the EA Community that she had installed the system but was curious about what types of prompts to use. And I think this question sums up the real problem: The technology is there, but nobody knows how to actually use it effectively. Almost all use cases aim towards doing things faster, chatting with files, renaming them, nothing we couldn't do before.
I hope to change that in these next tutorials (thanks, Debby!) and at least present some new ideas. In this week's tutorial, however, we start with the basics. Next week, we're gonna explore more in depth.
Summary: The tutorial answers fundamental questions about Claude code: What it is, why it differs from ChatGPT, how to use it for academic work, how to set it up locally, and how to load so-called skills to extend its capabilities. It is the first part of a series on Claude Code for Academics.
I also argue that the biggest problem right now is not technology itself, but a lack of clarity about how we can practically use it (most online examples tell you how to use it, not what to do with it, and most are quite trivial).