Some thoughts after using Claude Code in the Cloud for a project..

  • Design your (sub)system first, then discuss with AI. Don’t start with your thinking already pushed in a direction by AI.
  • Your names, not AI names (tables, fields, modules/classes, methods). You keep the mental model. You make sure code is in the right place. The best way I’ve found to do this is to concisely discuss the modules and methods with the agent before it builds.
  • Get as much out of the chat as you can. Eg, testing instructions into PR description and story breakdown in text file in the repo.
  • Do the work breakdown in vertical slices that add business value. AI is not good at this breakdown.
  • Every piece matters. Sure some are more critical (eg, money, personal data) but no part of your software is not important enough to do well. If it is that unimportant, you shouldn’t have built it.
  • Just because you can build a feature doesn’t mean you should. Do many of your users really need it? Does it really belong in the product? Would the users rather the software was faster and more reliable?
  • Read the integration tests. They should make sense and test the real thing.
  • You don’t have to use all your tokens. And chill, AI will wait for you without complaint.
  • If you are moving faster than you can mentally process, either slow down, or find a new way to interact with the AI to do the work with less mental burden and more clarity (and less rework).