Storing the plan and discussion as Markdown in Git is an interesting approach. It basically treats the agent’s reasoning as part of the project history, not just the final code.
Curious if others here are doing something similar to keep context across sessions.
> The workflow I’m going to describe has one core principle: never let Claude write code until you’ve reviewed and approved a written plan. This separation of planning and execution is the single most important thing I do. It prevents wasted effort, keeps me in control of architecture decisions, and produces significantly better results with minimal token usage than jumping straight to code.
One thing I’ve learned: if you notice the agent spinning its wheels, then throw the work away, identify a fix (usually to Claude.md) and start over. Context rot is real.
If you had a coworker that had tendency to hold everything about the project in their head you would push them to write stuff down eventually.
In strong words probably.
I agree with the other commenter who said they don't build anything without a plan. I would double down on that and say that you need to overplan, and regularly toss out plans as you do research/discovery.