An AI Chief of Staff Running on Your Own Machine
Published on 13.04.2026
I Built Myself an AI Chief of Staff. Now You Can Too.
TLDR: The author built "Claudia," a local AI assistant that runs in your terminal, maintains a personal database of conversations, tracks relationships and commitments, and improves with continued use. The article walks through the installation process and the philosophy behind context-aware AI assistants.
Summary: The piece opens with a bold claim: relying on your brain for memory is becoming obsolete, much like mental arithmetic was before calculators. The author argues that the real productivity gain comes from offloading the grunt work so you can focus on the actual problems. Claudia embodies this idea — she runs locally on your machine, stores every conversation, and builds a knowledge graph of your clients, commitments, and relationships. The author's daughter even designed the character, a small robot with yellow hair and a blue dress that sits 3D-printed on his desk.
The distinction drawn between the "R2-D2 model" of AI (execute a command, get a result) and the "Yoda model" (an entity that knows your context and challenges your thinking) is worth sitting with. Claudia has roughly forty skills — morning briefs, meeting prep, relationship mapping — and connects to email, calendar, and transcription tools. After seventeen hundred memories, the author claims she knows his world better than a human assistant ever did.
The article then shifts into a sales pitch: a $49 recorded install kit with a private WhatsApp group, and a $99 live workshop capped at ten seats. The underlying technical argument is that connectors — email, calendar, transcription — are where the real value lives, not in the base installation. The author's golden rule for getting unstuck is straightforward: copy the error, paste it into any AI, and ask for a step-by-step fix.
I'll admit, the Yoda-versus-R2-D2 framing captures something real about where AI tooling is heading. But I keep thinking about what the author conveniently sidesteps. Running an AI assistant with persistent local memory and full access to your email and calendar means that data lives on your machine in some database format. What happens when that machine gets compromised? There is no discussion of security boundaries, data encryption at rest, or what occurs if someone gains physical access to that laptop. The author calls it "not dystopian," but running a persistent memory system with access to your entire professional life does introduce real risk.
The piece also glosses over the fact that Claudia is built on top of Claude Pro, meaning you are still sending data to Anthropic's servers. The "local" part is the database and the terminal interface, not the language model itself. This is not a privacy play, despite the framing. And the $49 to $99 price tags for what amounts to running four commands and connecting a few OAuth integrations feel steep when the underlying technology is just Claude Code with a personality layer and a SQLite database.
Key takeaways:
- Local AI assistants with persistent memory and tool connectors represent a meaningful shift from browser-based chat interfaces
- The real value compound is the data accumulation over time, not the initial setup
- Error-driven learning with AI assistance is a practical onboarding strategy for terminal tools
- The security and privacy implications of persistent local AI memory deserve more attention
Why do I care: As someone who spends a lot of time thinking about developer tooling, the idea of an AI assistant with institutional memory about my projects is genuinely appealing. But I am not going to hand over my email and calendar to a local database without understanding the attack surface. The connectors are where the magic happens, and they are also where the danger lives. I would want to see encryption, access controls, and a clear data model before I commit to this.