Here’s a real-world, working example. Hope it helps someone:
Asked it to design a file spec for objects a carpenter would build.
Then asked it to make a sample file of a bench.
Then asked it to write a script to draw a blueprint from the file it wrote (in the spec it wrote, so wild)
It worked out of the gate, but wasn’t exactly right. We iterated on it a couple of times.
When script was nice I asked it to save it to it’s disk. This get’s it to do a final review of it’s own code and you generally get a working file.
Then asked it to make a test file that is a chair.
Then asked it to run the script with the chair file and it returned a ‘blueprint’ (haha) of the chair.
You then ask it to write up ultra-concise documentation for use by Gpt4. Download the script (or scripts if it’s more complex) and doc file.
Make a gpt, upload the scripts to it and put the instructions in the definition and you’re good to go.
This works in any ChatGpt with Code Interpreter. With this approach you can do quite a lot. I use it for things like taking notes, midi files, general images and so on, and the tools to work with them.
(shameless plug, my personal code writer: https://chat.openai.com/g/g-b8VNoIfxI-gptcode)