I would recommend doing this
Defining a function in the API request that tells ChatGPT to only return .json for a Node-RED flow
Then adding the .json flow into a separate tab so the user can inspect what ChatGPT generated before deploying
For simple flows ChatGPT might do OK
For more complex flows I think it would be asking too much to have it return perfection
But it would still probably return something that saves a bunch of clicks
Function calling:
Another idea - once I explain my flow requirements, ChatGPT could recommend the best modules/npm libraries to use and I could install them more easily. usually its a headache to hunt down the extra nodes needed.
And on top of that, Im sure you guys have tons of good example flows if ChatGPT got some fine-tuning based on prompt and flow examples, it would definitely get better at this task
Would be glad to collaborate on these ideas, I have been spending tons of time on ChatGPT in the last months