Claude Code drift tied to vague plans; CLAUDE.md fix urged

The problem is one of timing: once a coding agent is already writing code, any ambiguity left in the plan becomes a judgment call made with only the narrow context it has loaded at that spot. That is why it calls for a separate planning phase that forces the agent to read the code it will touch, check how mainstream tools and libraries already solve the problem, and test behaviors that documentation cannot settle.

It also treats the plan as a hypothesis rather than the source of truth, telling the agent to keep checking the plan against the actual code during implementation so it does not stop reading once the plan is written. Placing changes at chokepoints keeps one change from needing to be replicated across branch points, and adding unrequested features or opportunistic refactoring should be avoided.

A recurring theme is authority: if a revision cycle goes past two rounds, the agent should rewrite the whole plan instead of appending to it. The final gate is user approval, and if the approach changes mid-implementation, the agent is told to seek approval again at that point.