Recursive self-improvement just shipped as a slash command. Claude Code can now build an evaluation for your app, grade its own grading, then climb the score one change at a time against a held-out test set it never gets to see, so it can't cheat by memorizing the answers. Anthropic pointed it at Claude itself and watched it go from 66% to 88%. Here's what happened in that run. They aimed it at the skill that teaches Claude to write code against their own API. Claude added missing sections,…