“Run the same prompt twice, and once you will get reasonable output, and the second time it will fail to successfully run an Action or process the Action’s results.”
Macpherson wired his open-source tenforty tax library into a custom GPT and called it Tax Driver. It answered almost any scenario, and it missed the point of a stock-exercise question “in that cheerfully confident way GPT-4 has.” It claimed to draw a chart it never drew. A tax advisor that is right half the time is a liability with a chat window.