The curse of instructions is a named research finding that a language model’s ability to follow each individual instruction drops as more instructions are combined into a single prompt. Research on the effect reportedly found that even GPT-4 and Claude struggle to satisfy many requirements simultaneously: presented with ten detailed rules, a model may follow the first few and start overlooking the rest.
Usage
The finding is cited as the reason to decompose a large specification into sequential, simple instructions rather than a single prompt listing everything at once — focusing a model on one sub-problem at a time, completing it, and then moving to the next, rather than presenting the full requirement set together.