Antigravity: An AI Agent That Lives Inside Your Obsidian Vault

What happens when you give Claude Code direct read-write access to your Obsidian vault — and how a three-layer structure of Rules, Skills, and Workflows makes it actually reliable.

For years, I had a stack of things I called “work I should do but don’t.” Organizing old notes. Finding connections between articles I’d written. Checking analytics and thinking about what to improve. All of it valuable. None of it happening.

The bottleneck wasn’t time. It was friction. The work was unglamorous, repetitive, and interrupting enough that I’d avoid it until it became urgent — and then do it badly.

Antigravity changed that. Not by making me faster. By taking the unglamorous work off my list entirely.

What Antigravity actually is

Antigravity is a configuration for running Claude Code directly inside an Obsidian vault on your local machine. The key difference from a normal AI chat tool: Claude gets read and write access to your file system.

Most AI tools can only see what you paste into them. Antigravity can navigate your vault — read notes, follow links, append to daily notes, rewrite frontmatter, run searches across thousands of files. It does this through the terminal, with your vault mounted and accessible.

The practical result: you can ask Claude to do things like “find all notes that link to this topic and check if the descriptions are consistent,” or “scan last month’s daily notes and extract recurring themes,” and it will actually do them — autonomously, not by giving you instructions for how to do it yourself.

The insight that makes it work: delegate what you don’t want to do

Early on I expected the main benefit to be speed. I’d get things done faster.

That’s not quite right. The real benefit is more specific: things I was actively avoiding started happening.

Publishing a newsletter used to involve a sequence of small tasks I found annoying — checking formatting, generating a cover image, adding related article links, writing a description. Each task was ten minutes. Together they were enough friction that I’d skip steps, or delay publishing, or write less.

When I defined that process as a Skill (/publish), Antigravity handles each step and asks for my approval before proceeding. I just write. The tedious parts disappear.

The most concrete example: I had 500 articles accumulated over years. I wanted them reorganized into atomic notes — one idea per note, properly linked. In theory, doable. In practice, never going to happen manually. With Antigravity, I ran the process over several sessions and ended up with about 1,800 atomic notes extracted from the archive. That reorganization fundamentally changed how I can navigate what I’ve written.

Rules, Skills, and Workflows

The reliability of Antigravity comes from three layers defined in the .claude/ directory inside the vault.

Rules are always-on guidelines. Things like: prefer Obsidian CLI for file operations, don’t create links to notes that don’t exist, use specific date formats. These run silently in the background and keep Claude’s behavior consistent across sessions.

Skills are reusable procedures. /publish handles article finalization. /create-permanent extracts an insight from a conversation and files it as an evergreen note. /weekly reads daily notes and produces a summary. Each skill is a Markdown file describing a sequence of steps. Once defined, you invoke it with a slash command and Claude works through it, pausing for approval at key decisions.

Workflows combine multiple Skills into larger routines. A monthly review workflow might run /weekly across four weeks, analyze the results, and produce a prioritized list of unfinished threads. You define it once and it’s yours indefinitely.

Your notes are what make the AI actually useful

The deeper I’ve gone with Antigravity, the more I’ve come to believe: the quality of an AI agent is limited by the quality of the context it has about you.

Generic instructions produce generic suggestions. But when Claude has access to notes about how you think, what you’ve tried and rejected, what your current projects are and why — it operates differently. The suggestions start to fit your situation instead of being plausible-but-generic.

This reframed why I take notes in the first place. Daily notes that once felt like journaling busywork now function as training data for my AI collaborator. The more you record, the more accurately Antigravity can anticipate your intent.

The corollary: an Antigravity setup on someone else’s vault would behave completely differently, because it would be shaped by completely different notes. That context is the part no one else can replicate. It’s the only part of the system that’s irreversibly yours.