Michael Leibovich: How I Built My AI Chief of Staff

Linkpost | In Tech
| 2 minute read |by Constantin Gonzalez
A Chibi-style cartoon illustration of Constantin as a smiling man with a beard wearing an orange jacket and teal shirt, standing next to a small white robot with blue accents and a yellow star on its chest. Both characters are giving thumbs up gestures. In the background, faded office elements including laptops, tablets, and checkmark icons are visible, suggesting a technology or productivity theme.

Detailed write-up by Michael Leibovich on how he used Claude Code (inside the Claude Desktop app) for implementing his own Chief of staff setup.

As someone running a very similar setup myself, I particularly liked his use-case descriptions and the insights coming out of “externalizing your thinking”. I can totally relate to that!

Also, I agree that installing Obsidian and giving Claude access to it through its REST API plugin and MCP are great for giving it a structured, extra memory layer beyond its own memory mechanism. Claude can write Obsidian-style Markdown documents directly to the file system which is faster than going through the MCP/API, but the real value is the ability to search for existing documents through MCP that allows Claude to discover notes, make cross-correlations and more.

I would add a few things from my own perspective here:

  • You can use Claude Cowork (inside the same Claude Desktop app) instead of Claude Code. It’s the same underlying engine, but more friendly towards non-coders.
  • The Claude Productivity plugin helps you set this up, no need to copy/paste detailed prompts and documents.
  • The standard “Claude” way of providing global instructions is through a CLAUDE.md file. When starting up a new Cowork task, it sets one up by default, and reads it during every session. No need to write a GLOBAL_INSTRUCTIONS.md separately. You can also literally tell Claude to “update your CLAUDE.md file with…”.
  • Adding a joint task/todo-management tool (like Trello in my case) can also help a lot, beyond Claude Cowork’s built-in task-tracking (which is more short-lived and not intended to be used by its user).

I wasn’t expecting so much value out of storing call transcripts and can see now how it can be useful. Some people I work with use Google’s transcript feature for Meet or Fathom, though I prefer Michael’s local (and non-intrusive) approach to use Earmark. I’ll try experimenting with that, though I find it awkward to ask people in a call whether it’s fine recording it all the time. Perhaps it’s less of an issue and people will learn over time to be fine with it?

Visit Original Article →