Dev Rules
For humans working with AI on code. Non-negotiable basics.
- Open the correct folder. Not the one you think is right. The actual project root.
- Tell the AI your environment. Windows, WSL, Ubuntu, or whatever shell you’re in — they are not the same and the AI is not clairvoyant.
- Don’t change twelve things at once. If you changed twelve things and one broke it, you own that. Small, testable changes.
- Small commits. One logical change per commit. Your future self will thank you. So will the AI helping you debug it.
- Paste the exact error. Full message, full stack trace. “It says there’s an error” helps nobody.
- Say what you already tried. Prevents the AI from sending you in circles.
- Say whether it’s greenfield, maintenance, or refactor. These are different jobs. The agent will make different assumptions about risk tolerance, scope, and how much to touch.
- Say whether your local folder is the truth. If active work lives in a branch, or the local copy is behind remote, say so — otherwise the agent is working from a lie.
- Say whether you want a plan first or direct execution. Some people want a proposal to review. Others want the thing done. Neither is wrong, but silence gets you one at random.
- Say what the output needs to be. Local experiment, something you’ll commit, or something going live — those have very different standards of care.
Sources: ChatGPT (GPT-5.4), Claude Code