[dudeinhawaii]

I had the same exact experience with Bard/GPT3.5 and I'm using GPT-4 for a similar purpose of generating "analysis reports" from raw data coming out of a trading algorithm system.

GPT 3.5 had similar problems with formatting and following instructions, and I found that I had to move key instructions to the front of the description or else it would wander off the reservation.

GPT-4 operates as if I had dumped a collection of data on a junior analyst's desk and asked them to write up a report for our clients. It can even follow instructions like "insert market related joke or jab at the Fed here" in a completely believable way.