Three weeks ago I decided to build my own personal AI assistant. Openclaw was an interesting first go, but when the AI started talking to itself and other things just started breaking, I threw it all out for the second time, and then started to build my own AI from scratch with Claude Code. SO MUCH BETTER.
I don't have "10 tips for building your own AI from scratch with Claude Code" because I don't know more about vibe coding than any other AI-fizzing joker on the internet.
All I know is it is FUN. I did it roughly like this:
I don't have "10 tips for building your own AI from scratch with Claude Code" because I don't know more about vibe coding than any other AI-fizzing joker on the internet.
All I know is it is FUN. I did it roughly like this:
- Make it talk. A simple pipe between WhatsApp-Docker-Claude; send a message get a response.
- Make it understand my life (notes). The ability for the thing to query my Obsidian vault, and write things to it. It can read/write the actual .md files, but it also has the entire thing in a vector db.
- Make it remember. Added the ability for it to have it's own memory alongside everything in my "memory" (notes) so it remembers things about how I like to interact with it.
- Make it proactive. Added a 'heartbeat' so it'll review my notes, review it's memories, review conversation logs every few hours and form a perspective on whether I need to do/think/write something or take a real world action on something.
- Make it pretty. After a while, talking in WhatsApp on desktop started to annoy me so I built a prettier front-end for it on my Mac:
Pretty proud!