@cyrilXBT - Hermes Agent, Obsidian

Open original

@cyrilXBT on X The post lays out a three-layer setup for running a one-person company: Obsidian stores everything you capture, Claude Code processes and links that material, and Hermes Agent runs tasks in the background and gets better at them over time. It's a practical guide to reducing the digging and repetitive work that eats a solo founder's day.

  • Obsidian acts as the memory layer, with a raw folder for unprocessed input, a wiki folder for linked permanent notes, and a CLAUDE.md file that tells Claude Code how to handle the vault.
  • Claude Code reads the vault and turns raw material into connected notes, so each new source makes the next one easier to process instead of being summarized and forgotten.
  • Hermes Agent runs unattended, builds reusable skills from completed tasks, and works inside Telegram, so routine work like inbox triage and follow-ups gets handled overnight.

The core pitch is a three-layer stack for running a one-person company. Obsidian is the memory: a vault with a raw folder for unprocessed captures (articles, transcripts, voice memo notes), a wiki folder for clean linked knowledge, and a CLAUDE.md file at the root that tells Claude Code how the vault works. Claude Code is the processor. Pointed at the vault, it reads what you drop into raw, pulls out the new ideas, checks them against existing notes, and either extends or creates linked notes. The author's argument is that this compounds: the more you feed it, the more connections it surfaces, and it handles knowledge capture, decision logging, and drafting emails or proposals in your own voice. Hermes Agent is the operator, an open-source framework (the author points to MaxHermes, which runs inside Telegram and is set up at agent.minimax.io with no server or API keys) that works unattended and builds reusable skills from tasks it completes, so repeated work gets faster over time. The reasoning for why this is feasible now is that earlier tools each handled one piece in isolation, while Claude Code can now reliably act on a whole vault's context and agent frameworks can run for long stretches while learning from what they did. The author frames the payoff in hours returned: tasks handed off before bed, first-response customer replies overnight, routine research compiled while you sleep. The piece is promotional in places. It's a pitch for a specific setup, and the claims about compounding value and self-improving agents are asserted more than demonstrated, with no metrics, timelines, or failure cases. The practical architecture, separating raw input from curated output and making the processing step automatic, is the part worth saving. The product-specific details are worth verifying before you commit to them.

Ask about this article