I was reading people’s comments on a blog post about how they prompt an LLM. Someone said, “Today I don’t even think about how to prompt. I let AI help me prompt.” On one hand, the idea of meta prompting is interesting as a way to test the boundaries of human-AI collaboration. However, I’ve found myself writing more fragmented prompts lately.
It seems that as models get more powerful, I’m less inclined to organize my thoughts by writing a clear prompt. There’s no longer a big difference between what’s in my head and what ends up in the prompt field. Those thoughts would have gone through another filter layer (my own language-processing system), but now they land in my Claude chat as fragmented as they were when they first came to me. I am confident the model would infer what I mean from the bits and pieces, and even sort them for me. It feels as if part of my language-processing system is being lent out to the LLM. The time it once took me to flatten my convoluted thoughts has now become tokens used by the LLM.