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.