My Agent Setup and the Practices Behind It

Good morning!

I recently had a quite successful talk at O’Reilly where I shared my current setup for working with agents. I am offering the same talk (and way more) in our recently announced mentorship.

Here it is as well for all my paid subscribers since I promised earlier this year I’d always share my talks with them. Even if you are a free subscriber, I’ll still share some key aspects of my setup here!

If your agent has to relearn your preferences every time you switch tools, projects, or computers, your context lives in the wrong place.

The main idea behind my setup is simple: keep durable context outside the model.

My notes and personal skills live together in Obsidian. Shared project skills live beside the code in GitHub. Claude Code and Codex read the same plain Markdown instructions (skills), and every run records feedback (both on every use and on a scheduled regular weekly basis).

A longer prompt does not make an agent improve. A feedback loop does. Day 30 should be better than day one, and all your skills should have one.

I open-sourced a clean template version of the vault for everyone:

Clone the Obsidian Agent Vault Template

It includes agent rules, skill routing, a worked skill and feedback example, and templates for building your own.

I also recorded a private 15-minute walkthrough of the rest of my setup: multi-device sync, always-on agents, scheduled agents, scoped permissions, self-cleaning, my model preferences and how I combine Claude Code with Codex efficiently.

If you want our team to help apply this to your own production stack, the Towards AI Mentorship includes this blueprint, live Q&A twice a week, async engineering support, project reviews, and a new production blueprint every month. Otherwise, to access the talk and all my upcoming talks, you can support me by upgrading your subscription.

The full walkthrough…