GPTo1-preview + CursorAI: From idea to App

I've been playing around with CursorAI for a few weeks, and I just unlocked a super power: use o1-preview to write an entire development plan for you that you can then incrementally paste into Cursor's Composer feature.

-- -- --

Sections for this video:
1. Intro
2. Setting up dev environment with GPT 4o
3. Creating the dev plan with GPT o1-preview
4. Building the app with Cursor AI
5. Outro

This is a raw take. I run into bugs and fix them along the way. I hope this helps give you some ideas while setting realistic expectations around how incorporating these tools into your development workflow will help your process.

[arunkumar-uotg]

Can you please share the o1-prompt in the description as a word file or something. It's very neatly done

[Sidane]

This is excellent, thank you for walking through that in such detail. I’m going to give this approach a try ❤

[rtwg]

Super helpful to a n00b /wannabe dev, like me!

[cybersamurai]

Debugging code that is written by AI can be 10x times more complex than usually it would be. GPT sometimes would write 6 lines of code or complicate parses when it could be achieve much simpler. Its a trully hit or miss

[micuentadecasa]

Great work. Some of the things that you did should be automated, like the creation of new files, and I think that cursor should check the logs and propose changes to the errors, it has no sense we have to copy paste.

[davek.]

Thanks for sharing this! Very insightful approach leveraging the 01-preview model in the upfront design/dev plan, before leveraging cursor ai’s composer. I might have missed this part, but did you happen to try multiple cursor models, or are you most comfortable with Claude 3.5 sonnet’s output? Lastly are you using cursor pro subscription?

[MissplaySpotter]

As for some reason he doesnt share the document, here is it for you guys. (Might not be 100% 1to1 but close enough I guess):

"
Here is the complete content:

---

**I want you to create a development plan for me.** I am using Cursor, the AI powered IDE, and I will use the dev plan to prompt Cursor to create an app for me.

**My tech stack:**
- react
- MUI (including icons)
- express
- node
- axios
- dotenv
- openai api

**What I already have set up:** dev environment is setup with a basic front end and server.

**My app:**
- **title:** DebateGPT
- **purpose:** allow a user to generate a debate between two personas, about any topic

**Description of UX:**
There will be 4 inputs: debate topic, persona A, persona B, and argument count. Argument count defines how many arguments each persona makes in the debate—i.e. if argument count is 5 then each persona makes 5 arguments, totaling in 10 arguments for the debate. The user inputs necessary data and clicks a "generate debate" button. From there the debate begins to display in a chat-style layout on the screen. When the debate is done generating the user can download the debate as either JSON or Markdown.

**How the debate gets generated:**
When the user submits the form, pass the debate topic, persona A, and a meta-prompt over to openAI so it can generate the first argument. Immediately display the first argument in the UI for the user to see. Meanwhile, take the full context of the debate so far, the debate topic, and persona B, and send that over to openAI so it can generate the next argument. Repeat this process until the argument count is hit. Then stop generating arguments and display the download button.

**For the project plan, I want you to think like a product team** (dev lead, designer, product manager). Slice the work into small, focused scopes that can be reliably achieved. But don't make them so small that there are many more tasks than necessary. You want to balance the number of tasks you generate, with the quality and reliability of achieving those tasks.

**For the project plan, be sure to include 3 things for each scope of work:**
- a user story, describing user goals
- a design spec, describing ux / app behavior
- a technical spec, describing in natural language what should be coded, and in which files

**Here is the directory path for my current app, so you know which files I already have and where they are located:**
debategpt/
├── client/
│ ├── node_modules/
│ ├── public/
│ │ ├── favicon.ico
│ │ ├── index.html
│ │ ├── logo192.png
│ │ ├── logo512.png
│ │ ├── manifest.json
│ │ ├── robots.txt
│ └── src/
│ ├── App.css
│ ├── App.js
│ ├── App.test.js
│ ├── index.css
│ ├── index.js
│ ├── logo.svg
│ ├── reportWebVitals.js
│ └── setupTests.js
├── .gitignore
├── package-lock.json
├── package.json
├── README.md
├── node_modules/
├── debategpt.code-workspace
├── package-lock.json
└── server.js

---
"

[kamertonaudiophileplayer]

Is a way to reduce a natural language typing? For example: fix: <paste an error message here>. I noticed some people do it and it worked.

[kamertonaudiophileplayer]

ChatGPT, can you compose an entire installation script for all development components?

[KCMNJL]

I had to turn off Chrome Hardware Acceleration to watch this video. But only because I have a python script running which is fine-tuning GPT2-Medium with a modified network (added 24 Cross Attention Layers interspersed with the Self Attention Layers). The training run comprises of an audio embeddings file that I generated from 76GB of Ted Talk audio data, along with 123, 000 chunked and aligned transcripts for those audio segments. If that's largely gobbledegook to you.... I am essentially rebuilding GPT2 to have multimodal capability.... Voice Inference on top of Text Inference.

Am I some mad ML Nope.
I'm a software engineering generalist who specialises in nothing but getting passionate about problems I could solve if I had specialist info. In other words, I will never invent wheels, but I'll put all the existing wheels to great use.

Why is this all relevant to this video?

Well, before there is not a shot in hell I could have achieved any of the above with any of the wheels that existed.
It's so good, that I had an existential conversation with GPT-4o afterwards saying how I felt like a fraud and a cheat .... achieving a Training Loss graph that looked because I had just followed it's big brothers instructions.

But of course, it just congratulated me on my success.... and made me see that I'm not a cheat, I just had a vision and used what was in front of me to get it done.

Alright, enough about me....

Excellent video my man, have a sub buddy! :)