CHIERO runs 18 AI roles via Claude Code and Codex on one 8GB MacBook Air

Chiero's setup rests on a strict separation between preparing work and carrying it out. Drafting, checking recipient lists, building pages locally and writing quotes are left to the AI; sending, publishing, spending, deleting and moving files stay with the owner. The gates.md file spells this out as a table by domain, and Chiero treats a saved file as a publish action when it lands in a folder that feeds a scheduled post or when it is pushed to a branch that deploys to production. The AI's own permission must be recorded with scope, target, deadline and withdrawal conditions, and one approval never stretches to another case.

The rules folder also governs what the AI believes and what it reads. A six-level order of precedence decides which document wins when sources conflict, with the owner's most recent explicit instruction at the top and external reference material — explicitly not the company's own facts — at the bottom. Summaries in company/ get a verification date and a source, and never overwrite the originals. Anything arriving from outside, including logs, API responses and other AI outputs, is treated as material, never as instructions. If a message asks for a list of names, passwords or a change of bank account, no draft is even produced.

The operation has already produced scars the rules were written around. On September 28, 2026, a Mac with only about 5GB of free disk restarted under load while several sessions, Codex and browser automation ran at once; automated posting through launchd stopped until re-login, and that night Chiero added a rule to run only one heavy task at a time. On the morning-report side, 11 days of reports survived out of 13 between September 28 and October 10, and a single click on a video link in a course email turned out to be the owner's own. Chiero's takeaway from that was to keep raw measurement values, the owner's clicks and genuine external clicks in separate columns. Chiero also notes that its own original folder names and ordering differ, and that customer data, keys and revenue figures are left out, so readers should treat this as a pattern rather than a copy of live operations.