A second brain of plain markdown files, with Claude as the engine

I think a lot about how much of my time goes into capturing and tracking information. Knowledge work, basically. Whether I'm researching a project, organizing notes and action lists from meetings, or just trying to find the note I know I saved somewhere, I depend on pulling up the right information at the right moment to connect the dots and make a decision.

When knowledge capture is slow, or doesn't happen at all, the flow of information breaks down and progress stalls with it. My favorite example is Dr. John Snow tracing the 1854 London cholera outbreak to a single water pump on Broad Street, a story Edward Tufte retells in Visual Explanations: well-organized evidence, finally gathered in one place, pointed straight at the cause. My version of the problem is smaller. Transcripts pile up unread, action items scatter across a dozen meeting notes, and the note I need is always the one I can't find.

The way we do knowledge work is changing now that LLMs can understand, or at least usefully describe, the meaning of raw information. Hand one your meeting notes, a project brainstorm, or any unstructured file you can throw at it, and you quickly get back something organized and actionable. You show up already prepared to go to town on your action list, aware of what’s going on and ready to make decisions. This is the magic of inference, and Unicron is my attempt to build a daily workflow around it.

The architecture is deliberately thin: a front end on a folder of markdown files, with the Claude Code CLI as the engine.

The vault is just a folder: notes on people, areas, projects, and meetings, plus a journal, all plain text on the user’s own disk. Nothing is locked into a database or a single LLM, so the same folder opens in Obsidian, git, or any text editor. Keep it on Dropbox or iCloud Drive and the notes follow you across devices.

Claude does the work. It writes a daily brief from the whole vault with the top priorities, the threads you keep avoiding, and what’s due. Paste a raw meeting transcript and it becomes a clean note with an actionable plan, filed where it belongs. Schedule the brief and it runs each morning, even when the app is closed.

There’s no server. Unicron never collects anyone’s notes. When an AI feature runs, the content goes to Anthropic through the Claude CLI under the user’s own paid plan.

The app opens to Today and the daily brief. From there you get a Library of clean markdown, projects that hold months of threads together, every open task in one sorted list, and a graph of how the vault connects.

Capture is ⌘N. Add a meeting, task, project, journal entry, person, or area without breaking flow. Paste a transcript and Claude turns it into a clean note. Stray thoughts land in an inbox from a hotkey or a phone, and Unicron files them later.

The product decisions all trace back to one idea: respect the user’s files and their attention.

This is the natural evolution of the knowledge work I got interested in years ago, fresh out of an MLS program and working in digital libraries at Bell Labs.

My articles in Library Journal and LLRX back then were about the frontier of knowledge capture with grassroots tools like blogs and wikis, tools that gave access to anyone and empowered them to make informed decisions without being locked into cumbersome, esoteric systems. The Tools for Thought and Second Brain communities have been working this ground for years, and I’m excited to be building in it again.

Unicron started as Konigi Brain, a set of Claude skills running over my own vault. For a month I ran them from the terminal every day, and they wrote HTML files that were my morning brief and task list. They proved so useful that I spent a second month building the Mac front end around them: Tauri, React, and TypeScript, designed and developed with Claude Code. The landing page feels like the app itself: calm, white, rounded, one accent. Shipping it free-ish is deliberate. For now this is a fast way to iterate on the idea, not a revenue line.

You already see these ideas in Granola, Notion, Otter, Zoom, and Meet, and plenty of companies talk about pulling notes out of a dashboard and into where the work happens. Building my own end of it taught me a few things.