“An analysis of early user feedback for Claude revealed that 80% of complaints about "not following instructions" were due to the model's inability to correctly generate JSON output.”