I'm not a developer, and I run several Claude Code sessions in parallel. Rules that keep it from going off the rails
retained text
I've been building with Claude Code for a few months: a 1982 home computer emulator, a phone app, a few websites. I don't write code. What made it work wasn't prompts, it was rules: 1. Nothing goes live until I write "merge". Every session works on its own branch. 2. Each project has a HANDOFF file. Every session ends by updating it, so the next one starts where the last stopped. 3. Before anything counts as done, I ask for proof I can see: a test result, a screenshot, a live link. Not "it works". 4. Every background process gets a hard time limit. I learned that after an agent left a video render running for 9 hours. 5. One "manager" session watches the others and keeps a log of my decisions, and checks they were actually carried out. What rules have you ended up with?
document fetched 2026-10-30 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xigw3d/im_not_a_developer_and_i_run_several_claude_code