PPIG 2023 - 34th Annual Workshop
Prompt Programming for Large Language Models via Mixed Initiative Interaction in a GUI
Tanya A. Morris, Alan F. Blackwell
Abstract: Large Language Models (LLMs) now demonstrate many surprising capabilities that previously required special algorithms, for example interactive correction of syntax errors in structured text. However, the problem of how to systematically and reliably access these capabilities of LLMs has led to a new genre of “prompt programming” or “prompt engineering”. This paper presents a design case study in which we apply OpenAI’s Codex to an interface requiring syntax-constrained textual input, an email client. Via a mixed-initiative interface design, the system provides appropriate suggestions based on the output of the LLM. A user study was undertaken, finding that the incorporation of a LLM resulted in a decrease in perceived workload as well as a 62.5% reduction in errors. This work demonstrates how mixed-initiative interface design can better support attention investment in the use of LLMs, by delivering capabilities that might otherwise require prompt programming in a chat dialogue, but via a relatively conventional GUI.