My old full-stack case study

Nick AntonaccioAdmin
Apr 12, 2026 at 21:55 (edited, 3 revisions)

This old case study includes the full history I used to develop a complete full stack demo application with ChatGPT 4o. It's from an old generation of GPT with much smaller context size and much worse code generation capabilities (comparable to the open models which can now be run locally on inexpensive consumer GPUs).

Most importantly, it predates the zip file technique which I use as the cornerstone of project management now, but something in this study may still be useful, if only for historical context:

https://com-pute.com/nick/brython-tutorial-and-GPT-case-study.txt

EDIT: I update the project below - prefer to use the following links instead of the old one above.

History

Nick AntonaccioAdmin
Apr 27, 2026 at 04:33 (edited, 9 revisions)

In this case study, a fully functional CRUD app was created in 1 morning entirely by prompting chatGPT4o. This post contains updated links and relevant info from the original post.

A running copy of the application is available at:

http://cmpt1.com:5463

(username: joe@guitarz.org password: 12341234)

Here's a lightweight HTML version of the ChatGPT conversation that you can open quickly - this is the preferred link:

https://com-pute.com/nick/Case_Study_GPT_Full_Stack_Development.html

Here's the whole project - just the code (26k), without docs:

https://com-pute.com/nick/contact_app_no_docs.zip

Current setup instructions - as running at http://cmpt1.com:5463 - are available below (beware that original setup instructions in the zip file are deprecated):

https://com-pute.com/nick/case-study-gpt-contact_app--current-docs.txt

The application features include:

1) A complete authorization system built from the ground up, with validated email signup and automated forgotten password handling

2) A UI navigation system with collapsible left hamburger menus which show-hide responsively, according to screen size

3) An account management page which enables users to edit their own first name, last name, email address, and password (once logged in by the auth system)

4) A full-featured datagrid that enables inline editing of data values in a contacts database table (with full create, read, update, delete capability). There's no need to press any 'save' buttons while editing data in the grid - all edits are automatically updated in the database. Users can sort rows of the datagrid by clicking any column headers. Users only see and edit their own saved contact data (only data associated with their account, protected by the auth system).

All database schema, back-end logic, front-end logic, and every single bit of code was created from the ground up entirely by prompting chatGPT4o, and all debugging was performed completely by simply pasting error messages and questions into GPT. None of this application was created at all by any manual coding by a human, and every bit of functionality and styling exists *exactly as was specified in chat prompts.

History

Nick AntonaccioAdmin
Apr 07, 2026 at 19:23

Some additional versions of the conversation are available below. You shouldn't need these - they're included here for reference:

A .txt version which you'll need to download and open in a local text editor, in order for it to display properly (the browser link won't display all characters properly):

https://com-pute.com/nick/Case_Study_GPT_Full_Stack_Development.txt

A .mhtml copy of the conversation is available below - you'll need to download and open the local copy in your browser - it's slow to load, but the original formatting is displayed, and it can be stored locally on your own machine:

https://com-pute.com/nick/Case_Study_GPT_Full_Stack_Development.mhtml

The full original ChatGPT conversation can be found at the following link - beware that this original shared conversation can take a very long time to load and navigate with your browser (prefer the versions above):

https://chatgpt.com/share/c395ee46-40db-437e-b692-975652120f99

Here's a zip file with all the original code and the entire ChatGPT conversation in pdf and mhtml:

https://com-pute.com/nick/contact_app.zip

Nick AntonaccioAdmin
Apr 15, 2026 at 03:27 (edited, 1 revision)

One still valuable takeaway from this case study, is that even a model as incapable (by current standards) as GPT-4o, can eventually successfully build exactly specified software, if enough iteration steps are performed. GPT-4o was able to correct its own mistakes when enough debug errors were fed back to it, and reason just well enough to make progress.

So think of all the small models that can now be now run on expensive consumer GPUs, with approximately that same level of coding capability. Even if they can't produce the same quality code first shot, which GPT 5.4 (or the most recent Claude, Gemini, etc.) can produce immediately, the same sort of iterative process demonstrated in this case study, correcting error prone code in an iterative loop, can eventually lead to perfect output.

The big difference is that now anyone can hook up a little LLM to an agentic harness, and give it permission to perform all those iterations completely on its own, entirely hands off by human users, so that the iteration process can be performed much faster, fully unattended. The end effect is that you can get good quality end results from LLMs that perform far below frontier level.

History

Please login to post a reply.