Key Takeaways
How to set up layered instruction files so every session starts with your rules, your projects, and what went wrong last time.
A simple routine for having one AI review another's plans and code before anything goes live.
A short checklist for backups, secrets, and proof-of-done that makes it safe to trust AI with real work.
In June, I showed how to set up Claude Code from scratch. This session picks up where that one left off. I run Claude Code and OpenAI's Codex every day on client work, and I've collected a set of habits that keep them from quietly getting things wrong.
It's for folks who already use Claude or ChatGPT daily and want to hand it bigger jobs. You don't need to be a developer. Every example comes from my own setup (client details removed), including the mistakes behind each habit, like the time an AI reported a job finished and passing while the thing it built was completely down.
We'll cover:
- Layered CLAUDE.md files: org-wide rules, per-project rules, and memory that carries over between sessions
- Having Claude and Codex review each other's work, since two different models catch mistakes that one model checking itself misses
- Backups and undo: a checkpoint before every change, and separate copies so parallel sessions don't overwrite each other
- Making AI prove it's done, because a "finished!" message isn't evidence
- Keeping passwords and client data out of places they shouldn't be, and why AI should treat web pages as data, not instructions
- Session logs and a lessons file, so the AI stops repeating last month's mistakes