How to use GPTs to build a full stack app as a non-developer. Framework proposal

[dkampien]

Im thinking of a way to use GPTs to build a full stack app by myself as a non-programmer. I tought about a potential framework but before I test it out I want to share it here and get some opinions.

It’s obvious that you can’t build a full app in a single thread so the first step in this framework would be to split “the app” into multiple threads. Either by module, component, feature or problem. Assuming you know the process of building an app (ux, system design, architecture, code), the first thread could act like a discovery thread where you also find out what you don’t know, make a plan etc. And even if you don’t know, you include the learning as part of the process. You take those insights from the first thread and dump them into a centralized app documentation.

Then, you repeat the process for some other aspects, modules or components. And using the previous thread insights as INPUT for the new thread.

Now the question is how you feed the growing app documentation (could be a txt or pdf) into a new thread everytime. I know you can just copy paste as a prompt but I want to explore if GPTs could help in this scenario. You could:

1. Make a gpt for each app role (ex designer, front end dev, backend dev etc) with their own parts of the app documentation as knowledge bases. Then maybe you can mention them in a thread. (The mention GPT feature is what sparked this whole idea)
2. Make a single GPT for a software engineer and have the whole app documentation in a single (or multiple) file.

We know that the size of a knowledge base file is about 2 mil tokens. So its enough to cover an app documentation. Altough I haven’t thought about the actual code implementation and code base.

The only flaw im thinking right now is that GPTs retrieve knowledge via RAG (I think), and even if the app documentation has 2mil tokens, when retrieved to the actual chat it would still be limited by the model context window. Even if gpt 4 turbo has 128k context window its not holding all the context. Search youtube for “gpt 4 pressure testing”. And that was the whole point. For the model to know the app documenation as it grows bigger.

Can GPTs be somehow used in this context? Can this framework be refined to actually do something usefull in this context?

[RonaldGRuckus]

I can see where you’re going with this, and I think it’s a great thought.

There could be internal documentation with numerous LLMs that are assigned specific responsibilities, and they write their updates/progress to the documentation.

It would be a massive system, and I can easily see it being the future. Maybe once hallucinations and the attention loss for long context is fixed. Otherwise it will just lead to loose ends and incoherent code. BUT, I think at this point when websites (including the back-end) can be whipped up at a whim they will be something entirely different…

I don’t see it happening soon. It would probably be cheaper just to hire a full-stack engineer LOL.

I think a big issue with GPTs creating websites is the timeless problem that plagues any no-code solution for programming: User wants Thing A, it’s created haphazardly. User then wants Thing B, created haphazardly. Thing A was never meant for Thing B. User does not understand why it doesn’t work and bashes Thing A & B together until it works.

Then, the dreaded Thing C. That’s when the full-stack engineer gets called, and a therapist gets paid.

I mean, I can EASILY see that pretty soon we can have GPTs that can call eachother for expert opinions. I’m pretty confident that I’m going to have a Discord channel where different GPTs argue over trending news.

[mouimet]

Hey guys! it’s totally possible