Personal build · AI operating system
The AI operating system I run my work on.
A set of AI agents with clear jobs, an independent reviewer, and a hard line on anything that leaves my Mac.
Counted fact to come from project records
Counted fact to come from project records
Counted fact to come from project records
The problem
I juggle a job search, AI learning, a podcast and a household across a lot of apps. Single AI chats helped one task at a time. Then they forgot everything, and I couldn’t trust them to act on my behalf.
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer posuere erat a ante venenatis dapibus, posuere velit aliquet.
What I owned
The operating rules: who plans, who executes, what needs my yes, where status lives. Claude plans and writes the skills. Codex runs the scripts and scheduled jobs.
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer posuere erat a ante venenatis dapibus, posuere velit aliquet.
How it went
Split planning from doing
One model plans and writes the spec. Another executes it. Neither grades its own homework.
Review before anything runs
Every plan, skill, new schedule and anything that sends gets an independent review by a second model first. I see the objections before I say go.
An ask-first line
Email, texts, calendar edits, purchases and deletes need my yes in the chat that does them. A relayed message carries context, never authority.
Receipts over memory
Dated reports, an append-only log and a decisions log, so any chat can pick up where another left off.
How I know it works
- Placeholder: Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer posuere erat a ante venenatis dapibus, posuere velit aliquet.
What I’d do differently
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Integer posuere erat a ante venenatis dapibus, posuere velit aliquet.