[Terretta]

These are generative models where the LLM prompting works by examples it tries to match while generating new tokens into the provided pattern.

Chat implementations provide your recent back and forth history as example to the next prompt. This is how providing an example reply in JSON can get it to reply with new data using JSON.

Ongoing replies reference what it interprets as labels (the blanks to fill in) provided in the examples, replacing those tokens with new tokens.

Look at your first question and second question again. Pretend the first reply is in JSON but with placeholders for song title, artists, topic, lyrical analysis, and critique.

Pretend the second reply was supposed to use the first reply as its example, and replace only new data.

You see it put new generation into the provided example structure.

Now you can see what it did there. It did exactly what you prompted it to do.