4 ms·
I have done some experiments with telling ChatGPT it's an AI software engineer writing a Node.js program and it's talking to a special command interpreter that
by ilaksh 4y ago
I have done some experiments with telling ChatGPT it's an AI software engineer writing a Node.js program and it's talking to a special command interpreter that understands things like INSTALL_PACKAGE and UPDATE_FILE, then executing it's commands automatically and feeding the output back in. It was fascinating watching it try to complete a programming task, searching for appropriate modules, reading the docs, etc. But it seemed to get a bit overwhelmed and started alternating between listing files and editing the devnotes.md with it's plan. For some really simple tasks though it was able to complete them fully autonomously.
What actually worked better was just making a list of prompts and filenames in a JSON file and automatically giving each one to it and saving the results one after another. Then improving those specs after looking at the output. By being very specific, making sure that things like names and endpoints matched in each prompt, and adding a few hints, I was able to get it to complete a full working program with multiple HTML/JS pages and an Express.js server script.
Next experiment is to get it to output the spec.json based on another less technical one that doesn't have the file breakdown or technical details.
- idle_zealot 4y agoI've had success by asking it to generate project directory listings and then prompting it to show me the contents of each file. If the response token limit were removed I could probably compress project generation into a single prompt this way.
- ilaksh 4y agoRight, something like that is the next step for my automation program. The token limit is something I am really really hoping they will have a workaround for when they put out an API. Assuming they do release an API at some point for ChatGPT. If there was just an option to say "please continue where you left off seamlessly" that is what I really need. But may not be possible. Anyway I have been trying to avoid getting cut off by always telling it to ensure each source file is 60 lines or less. But the main thing is just breaking everything up into pieces.
- idle_zealot 4y agoReplying with "continue" does basically what you want. The only issue is that it breaks the webpage's formatting when it tries to continue in the middle of a code block. It will produce the rest of the code but it won't be in a code block and some special characters get interpreted as other format codes.
- sonovice 4y agoI never had any problems with "continue your last output".
- jvm___ 4y agoYou can talk it into generating rails code, and then ask for a react front end