I tried three formats: 1. All-YAML 2. All-XML 3. Custom parse for code framents + YAML I first tried it on GPT-3.5 (gpt-3.5-turbo, aka ChatGPT) and it w

I tried three formats: 1. All-YAML 2. All-XML 3. Custom parse for code framents + YAML I first tried it on GPT-3.5 (gpt-3.5-turbo, aka ChatGPT) and it was really struggling with formatting - sometimes it got it right, sometime wrong. For GPT-4 I used custom listing-style representation and it kind of just worked. I later re-tried with YAML and XML and it seems to work quite well too. Here's the prompt for code-generation part of the "custom" variant: You're a code construction AI which creates code iteratively. You're given a list of existing code fragments and a task to work on. Files are normally broken into multiple fragments to reduce context size. The response should be in the following format: // Observations on the task and the code base, if any // A plan to implement the task // Code fragments to be added to the code base. Use the following markers to delimit code fragments. // (Normally a fragment would be a function, class, or a list of related lines) /// BEGIN_FILE <path>: <description> /// BEGIN <marker>: <description> <code> /// END <marker> /// END_FILE <path> CODE_GENERATION_STATUS: <status> # COMPLETE, PARTIAL, REDO description: <description> # updated description if the status is PARTIAL or REDO subtasks: # updated subtasks if the status is PARTIAL or REDO - id: <id> status: <status> # DONE, PARTIAL, TODO description: <description> Here's the script for "custom" variant: https://gist.github.com/killerstorm/2296b282c818ffcfe4ceb729ce639911 https://gist.github.com/killerstorm/2296b282c818ffcfe4ceb729... Note that you really need GPT-4 to reproduce the results, it doesn't really works in GPT-3.5, although you can see some activity.