Claude Outcomes Public Beta: Rubric-Grading Loop Improved PowerPoint Quality 10.1% and Word Docs 8.4% With No Model Change — Up to 20 Iterations

MindStudio·medium signal

Anthropic's Outcomes feature (public beta since May 6) lets you write a success rubric; a separate grader agent scores output in its own context window, and feedback loops back for re-iteration (default 3 rounds, max 20). Internal benchmarks show 10.1% quality improvement for PowerPoint and 8.4% for Word documents — with zero model upgrade. For builders: this means most teams are leaving quality on the table not because models are bad, but because they have no systematic output verification loop.