A team at Shanghai AI Lab says most AI errors stem not from bad models but from thin prompts. Their solution--"context engineering"--shows that giving language models richer background information leads to better results.