Here’s how I decide which steps of a task to defer to an LLM:

- If the answer to the some input must be the same every time, then an LLM is probably not the best option, or at the very least, the output should be reviewed by a human

- If there is wiggle room and multiple acceptable answers, then an LLM can perfectly perform that task

In many cases I use a combination of LLM, deterministic code, and human review at key points or end of the workflow.

It doesn’t make me 10x more productive, but it gives me 100% confidence in my work.