Comment by blourvim

You could have limited-instruction llms where the model does one thing, for example summaries. It could accept a limited amount of instructions for example, first token for verbosity, second for style etc...