Founder & CEO, PromoPilot
Five rules that keep my AI agents from breaking production
Most of PromoPilot's code is written with Claude Code, and I run it as a solo founder from Kyiv: a web app, iOS and Android apps, a Chrome extension, a public API and 100+ free tools. Writing the code was never the hard part. These are the rules that actually stopped incidents for us.
- A separate tester session checks every visible change. The session that built a feature doesn't get to mark it done. Another session opens the page like a new user, clicks through it and sends bugs straight back to the builder.
- Money and real users stay out of tests. Payments, bonuses and emails are tested only on dedicated test accounts, never on a real customer "just once".
- Scheduled jobs need a heartbeat. A hung run once silently blocked the next ones. Now every scheduled task records when it started, and anything stuck too long gets stopped before the next run.
- Decisions go to a board, not chat. When an agent needs me, it creates a card with buttons for the possible answers. Chat messages get buried, cards stay red until someone answers.
- Deploy only the files that changed. One deploy reset folder permissions and broke blog images until we noticed, so now uploads never touch directories.
None of this is clever, it's just what broke first.
What rule did you add after an agent broke something for you?






Comments
0 comments