Your exploration session is the worst one to implement the solution.
You've spent an hour exploring a codebase with Claude. You ask it to implement the fix—and it repeats mistakes, forgets what you discussed, or gets confused. LLMs anchor to their own outputs, repeat errors, and degrade at 20% context fill.
The fix: explore in one session, execute in a fresh one.
-- DISCOVER --