A few things that have worked well for me:

  - Like normal software engineering - building small parts in isolation, verifying they work and then bringing them together to make the whole.

  - Asking Claude to do things using sub-agents to keep the parent context clean. i.e. reviewing plans, research, code reviews
  
  - If there's ever something you can have verification on - a test suite, a screenshot with pixel reading, a certain output to match. Then with `/goal` it tends to work very well with little guidance (in my experience)
  
  - If it keeps getting things wrong, then I make it write a skill. i.e. for some reason it always had errors running playwright but would get there in the end, and it would always have trouble with auth on my NAS but with the skill it can follow the path that's worked for it in the past and not burn tokens