AI-in-SDLC

Musings on Generating Code with ChatGPT

and a desired alternative

(all words by the editor)

Challenges with ChatGPT

Workarounds

One way to combat this is to remember that when ‘conversing’ with Chat you are actually simply sending an abbreviated history of your prior conversation, mostly consisting of your queries, and very little of the assistant’s responses, and then getting an answer simply based on that historical context and your latest query. The “conversation” with the underlying model is like talking to a person with zero short term memory where every question has to repeat the context of said question – it can get mighty tedious and frustrating. Chat is merely an abstraction, an opinionated simplification of that tediousness.

You can work around the issues with Chat’s design by manipulating the context; flood the user query buffer but with a context useful to the assistant, such as

Desired Alternative

It would be helpful to have an alternate Chat interface for iterative, collaborative work such as code generation/fixing.


Commentary by GPT-4, along with further brainstorming: https://chat.openai.com/share/1171acc9-67dc-4445-bb89-da6b0c9063f8 — Back to Chapter 4