Due to recent news about updated rate-limits and people wondering how people hit them I thought I would share how I work with claude code. Unfortunately this is only linux only and designed to fit my workstyle, so this is not one-size-fits-all.

When working on massive plan.md files, this usually ends up taking 5-7 iterations each talking around 15-30 minutes. What I usually do is create a plan before going to sleep and then come back to see what it did via quick glaces at work-log.md and git commits. This allows me to undo things I don't like or I want to implement myself or push the agent into a different direction.

Then updating the plan with (Changes Needed) helps correct any mistakes the model did in the plan execution which overall allows me to produce a pretty good end-product.

All the commits the AI made are then squashed, all the changes are looked through and the final commit is formulated.