Teaching claude to write pony introduces the idea of iterating on the alignment doc for the AI.

The tl;dr have a  /.claude/CLAUDE.md which is continually updated by the user after each session. This allows a period of reflection on what went well/badly and to fix this up.