A guy at the AI meetup in Austin told me my prompts were doing too much, and I finally fixed my workflow
Last month I was at an AI builders meetup in Austin showing off my chatbot and a guy named Dev just stared at my prompt for a second and said "why are you cramming six jobs into one prompt, split that up." I had been running one giant prompt that did tone, ticket sorting, and a summary all at once and it kept dropping steps about 3 out of every 10 runs. I split it into a chain of four smaller prompts that each do one job and hand off to the next, and my fail rate dropped to almost nothing on the same test set. Turns out the model was choking on competing instructions, not the task itself. Anyone else here break their prompts into steps like this, or am I late to something obvious?