#Coding support
1 messages · Page 1 of 1 (latest)
Imagine that instead of GPT you had a junior programmer who "goes off and does stuff" and their work "drifts" because they didn't understand the task well enough and you didn't guide them enough to keep them on track. It's similar.
A lot of people seem to like "vibe coding". Personally I find it awful, so the trick is to do the planning and constrain the AI so that it only does what you want it to.
Or, you take on the core work and just get it to speed up the work by coding "leaf nodes" only.
Depends on what you’re using to code I guess. I don’t have much experience with codex, but when in a regular prompt window try breaking it down into different chats. The context window is only so long so the responses will get a little wonky the more work you do in a single chat. Try a new chat, input whatever code you’re working on and reference the issues you were facing in the prior chat. I find you get better responses
You can vibe code but you must break down the problems into very bite size operations. I use codex and you just need to monitor and review the code being generated. If you develop and whole ream of code and it is a surprise it works, you are heading down a hole. Let it take care of the repetitive easy code and bounce ideas of it, always plan with it, review and review and step through the code. It may be the fact that it takes the same amount of time, but you are exposing yourself to different coding styles and methods. LLMs are a productivity tool, only.
I agree with this approach, I too have tried with breaking my requirement into multiple small task and then ensemble them into one after validation.
is there a guide on using codex for complete beginners?
- I connected it to a Github repository.
- I have two different chats, one for design and one for execution. The design is going really well and I'm starting to create the back end.
- Codex wrote some code for the backend, but I have no idea how to deploy it or test it.
To test it you probably want to have an IDE and link the git to the repository and test in in in the IDE
5.3 has a plan mode. I've used it to great effect in vs code via the official extension. Basically you will sit down and outline goals for the coding, and it will ask a series of clarifying questions then present you with a coding plan. If you approve it it'll get started and tell you what it's doing as it goes and you can either queue up comments or send them immediately to steer how it's working.
You can't just tell it to make you an app and then go make tacos but you can totally vibe code and guide it actively and get stuff done quickly.
get yourself this audiobook
TLDR: I've found this flow below helps, copy pasta from chat does not!
You are going to have to get into it at least a little bit. Figure out the corner of tech that you are interested in expanding and building in -- once you identify that, learn a little bit about the high level concepts. Use chatgpt to generate some reports, as well -- and this is important -- teach you what you need to know to be a good project manager of what you want to do.
I recommend using Codex for vibe coding, not just copy pasting what spits out in the chatgpt window, since sometimes it may be speaking generally and not evironment specific, etc, etc.
Those might seem like small things, like does the code give an example variable somewhere. It's something a dev might notice when prompting in chatgpt window. But since you're new to programming, using Codex standalone can let you understand how project folders get set up, watch them build out at you chat, etc.
And if you've already asked in another chat thread to create a project plan/outline and put all this into a zip, you can just bundle and give that to codex. And since codex has your local computer folder workspace accessibile you can say, "here's this thing I'm planning to make, let's go step by step and build this. Guide me as a professional engineer and mentor."
TLDR: I've found this flow below helps, copy pasta from chat does not!
yes, progress as you are until you feel frustrated. The move to VScode, WSL and openai API. but you must learn about the having an Architceture.json, requireents.txt and a milestone.md to chunk the tasks that the api must complete.