Reflection

This is the skill that makes Claude learn from mistakes and user corrections. It kicks in when a tool fails or when you tell Claude to do something differently, analyzing what went wrong and proposing specific updates to either skill definitions or a CLAUDE.md preferences file. The key constraint is it only suggests one change at a time and waits for your approval before applying anything. In practice, this means fewer repeated mistakes and less need to give the same instructions every session. The focus on tool failures is smart since those are concrete signals something needs fixing. Worth noting it tries to balance being specific enough to address real issues while staying generic enough to apply broadly.