As someone who’s spent an unreasonable amount of time working with prompting for both text and text-to-image, I can confidently say prompt engineering does work

As someone who’s spent an unreasonable amount of time working with prompting for both text and text-to-image, I can confidently say prompt engineering does work for any nontrivial use cases (e.g generation with constraints). I don’t like that prompt engineering works, but it does, and proper use of it does result in objective improvements in generation quality. The latest ChatGPT model, gpt-3.5-turbo-0613 has better system prompt steerability, and with some prompt engineering I can get GPT-4 quality results out of it at a fraction of the cost. Ultrapopular tools like LangChain and AutoGPT are essentially just prompt engineering under the hood.