How can LLMs understand and perform meta tasks? (e.g. summarization)

I don't ask how to make it summarize xy but if it is known how a "LLM" understands and performs this meta task at all.

The same is true for prompts like "Be brief" or "Explain in a few sentences". Since this works, the LLM must actually have an understanding of text length.