How good is AI at getting to an MVP (Minimum Viable Product) stage?

What is an MVP?

Prototype - interactive mockups eg Figma..or code to validate concept and flow. Often doesn’t have backend logic

Proof of Concept - feasability of idea.. not User Experience or scale.

MVP - smallest thing that delivers value. Should work reliably enough. Avoids tech debt so massive that can’t iterate

Source on GH for this experiment

Lets assume

  • We know who the customers are
  • We are solving a problem for them
  • They want to ‘Give us their money’

What is the sample application

Lets also assume it will

  • Have a database
  • Be a web application
  • Have reasonably professional design
  • Be as simple to understand as possible by a developer

How much to auto-gen?

20+ years of experience of CRUD app builders, many languages and many different codebases (yes enterprise applications specifically) have left me with strong opinions

  • Huge layers of abstrations in case you need it, increases cognitive load
  • I prefer the simplest possible solution
  • Repeating yourself a few times is okay in a code base to reduce complexity
  • I prefer technology I can understand (no black magic - I’m looking at you Next.js)
  • Complexity always comes back to haunt you

So although I could do a super fast build using React (which I don’t know) and Express on the backend (which I don’t know) or Next.js everywhere… I’m not going to

This is an MVP so tech debt is important.

I also enjoy the artistic beauty of simplicity.

1. AI - start with spec.md

Initial ChatGPT conversation about AI development process

This was my start with this prompt:

you are an expert software developer. Describe in detail how to go through the process of using ChatGPT to rapidly develop a new proof of concept web application.

I will give AI a sample project structure which includes:

- python and fastAPI for the backend
- postgres database which I'll host on my own vm
- front end html/css and javascript
- front end includes a sample page with css grid for layout, tailwind 4 

what I want to focus on is the AI process ie

- develop md files with a project spec in it (or mdc files?)
- develop coding style guidelines for ai
- use browsertools mcp (with cursor) to test the javascript
- do screenshots of the app to help in testing
- setup .cursorrules file
- setup <xmltag>code here</xmltag> style to help AI understand what is code 
- a spec.md file
- what is an llms.txt file

Then next prompt all around spec.md ChatGPT conversation about developing spec.md

please help me develop a spec.md file to use in developing a mvp

the spec is around a project called TLDW - too long didn't write

go into great detail - I want to know each step of this spec.md process

<oldprompt>
in relation to this prompt:
you are an expert software developer. Describe in detail how to go through the process of using ChatGPT to rapidly develop a new proof of concept web application.

I will give AI a sample project structure which includes:

- python and fastAPI for the backend
- postgres database which I'll host on my own vm
- front end html/css and javascript
- front end includes a sample page with css grid for layout, tailwind 4 

what I want to focus on is the AI process ie

- develop md files with a project spec in it (or mdc files?)
- develop coding style guidelines for ai
- use browsertools mcp (with cursor) to test the javascript
- do screenshots of the app to help in testing
- setup .cursorrules file
- setup <xmltag>code here</xmltag> style to help AI understand what is code 
- a spec.md file
- what is an llms.txt file
</oldprompt>

Start with a sped.md file

Go through each section with an LLM iterating

  1. Project Overview
  2. Core Features Checklist (which I’ll use to get the AI tick off when done)
  3. User Flow - eg new user, returning user, ai assist, templates
  4. Technical stack
  5. Data Models
  6. API endpoints
  7. UI Layout

Interestingly I don’t have a good sample project yet for this full stack, only a partial FastAPI project. Let’s start with the and see what happens.

So I now have a spec.md file with a good amount of detail.

I have a working Python FastAPI backend working on my laptop with uvicorn as the webserver.

Next step is to get a frontend working.

Generate a front end based on @spec.md

It will generate tons of code, which may well work. But…

Options

  • Full server side rendered eg .NET.. I know I could do this, but want to explore a more modern way, and it will probably fit the use case better using an API

  • API’s - lets just try this as I want to.

Prototype (no db backend) - file structure

So, we’re doing a Prototype which comes before the PoC and MVP

alt text

vscode running uvicorn via ctrl F5. Uvicorn serving up index.html, and the api routes which are running FastAPI.

alt text

/docs are the Swagger docs, which are interactive ie I can run the route, and even post data.

API testing (no UX)

How to get data from the API and display on page

  • Button click on page
  • Return json from API
  • Render on page

alt text

Raw javascript getting data using fetch. Tailwind 4 CSS (with no bundling)

Update spec for offline feature

I iterated on this feature for the simplest possible thing to start with ie a message saying offine, then just saves when back online. No local storage or conflict resolution at this stage.

HERE TODO next

  • securing the API (see chatgpt latest conversation)
  • Look at Paddy’s codebase to see process layout of .md files
  • Patch in Typescript?
  • I wonder if I should use a framework here?
  • POSTing data

2. AI - start with engineering ie scaffold out

Below are notes on a more classic way of using AI to solve small engineering problems (which it is great at).

https://chatgpt.com/share/6889f8be-dd00-8006-b640-76e80c748d94

I typed in the requirements into ChatGPT which in itself is a great exercise.

I’m working with an experienced front end developer, so I’m going to focus on the back end part of this application and see where AI tools get me

  • Can it get me to an MVP fast?
  • Can the code be any good?
  • Can I use a different framework (eg FastAPI which looks reasonable) which I’ve never seen
  • Can I use a language (Python) which I’m only reasonably familiar with, but I know AI tools have a lot of knowledge about.

Notes so far

  • ChatGPT good at coming up with frameworks to consider
  • ChatGPT great a solving specific problems (eg engineering) - python versions, pyenv, poetry
  • Cursor (Claude) when I asked for a simple project for FastAPI gave me some outdated dependency management which I knew I didn’t like (pip install)

Python

As we’re using python lets do it properly, and use AI to help solve engineering / version challenges

As I know we’re going for a custom build of a VM (I’m good at this), lets find out the best versions with all latest speed improvements to install.

3.13.5 was released on 11th Jun 2025 (about 6 weeks ago)

I’m using 3.12.3 on another project, which seems to be the most stable and supported.

3.13.5 is latest stable with small speed improvements over 3.12. Optional JIT is experimental but promising.

Python 3.12.3 comes with Ubuntu 24.04.2 (Latest LTS as of 31st Jul 25 and 5 years of security updates) which I’m using on dev and prod. Ubuntu 25 (only 6 months of security updates) comes with python 3.13.3

pyenv for multiple version of python on a system

I’m going to try 3.13.5 but need to isolate it, as I use 3.12.3 for other projects.

https://github.com/pyenv/pyenv

# pyenv to have different versions of python
# maybe don't need on production but do on dev
curl -fsSL https://pyenv.run | bash

echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.bashrc
echo '[[ -d $PYENV_ROOT/bin ]] && export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.bashrc
echo 'eval "$(pyenv init - bash)"' >> ~/.bashrc

echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.profile
echo '[[ -d $PYENV_ROOT/bin ]] && export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.profile
echo 'eval "$(pyenv init - bash)"' >> ~/.profile

## restart shell
exec "$SHELL"

## 2.6.5 on 31st Jul 25
pyenv

# install python dependencies before attempting to install new python version
sudo apt update; sudo apt install make build-essential libssl-dev zlib1g-dev \
libbz2-dev libreadline-dev libsqlite3-dev curl git \
libncursesw5-dev xz-utils tk-dev libxml2-dev libxmlsec1-dev libffi-dev liblzma-dev

# list available versions
# there is a t version, which means tuned. don't use yet.
# pyenv install -l

# this downloads and compiles the python version
#pyenv install 3.12.3
pyenv install 3.13.5

# should see 3 including system
pyenv versions

cd code/tldw
# I want to keep 3.12.3 as my default, but I found global system had shim errors when running my test python program so best to use 3.12.3
# pyenv global system
pyenv global 3.12.3

# this writes a file - .python-version in the code/tldw directory specifically for this directory
pyenv local 3.13.5
# pyenv shell 3.13.5

poetry install
# if big changes from previous
poetry lock
poetry install

# to see the path to python
# then select interpreter on vscode/cursor
poetry env info

Project Dependencies

So immediately with prior knowledge I know there are good ways of dependency management in python in 2025

  • poetry (latest version is 2.1.3 on 31st July 2025)
  • hatch

chatgpt seems better at creating an init

# initialize in an existing folder
poetry init

# Add dependencies
poetry add fastapi uvicorn

# I need to create a file here

# Run the app
# it was tldw_app.main
# but a mix of tldw-app folder, and tldw_app class... don't want.. so try none
poetry run uvicorn tldw.main:app --reload
# or if in the tldw folder
poetry run uvicorn main:app --reload

alt text

A test app making sure I have cursor working with the correct version of python.

FastAPI

As I know what I’m doing (roughly), lets see if the docs can get me going faster than AI

“FastAPI is a modern, fast (high-performance), web framework for building APIs with Python based on standard Python type hints”

https://fastapi.tiangolo.com/

# this wont work - need fastapi[standard] extras
fastapi dev main.py

# an ASGI server - what runs the FastAPI app and serves on http://127.0.0.1:8000/
uvicorn main:app --reload

# prod
# looks like could use gunicorn to handle multiple workers, but raw okay if managed using systemd and small/medium apps
uvicorn main:app --host 0.0.0.0 --port 8000

alt text

http://127.0.0.1:8000/docs - interactive Swagger UI which lets you test the endpoints.

Also http://127.0.0.1:8000/redoc an alternative, but not interactive ie can’t test. Can generate an OpenAPI.json file http://127.0.0.1:8000/openapi.json which can be used by Postman or code generators.

Okay now I understand what is going on, lets get AI to generate a new endpoing. It does it very well! /items/foo

# FastAPI class inherts from Starlette which is an ASGI framework for building async web services in Python
from fastapi import FastAPI

# create an instance of the FastAPI class
app = FastAPI()

# path operation decorator
# path is also called an endpoint or route

# operation refers to one of the http methods eg get, post, put, delete
@app.get("/")
# path operation function

# https://fastapi.tiangolo.com/tutorial/first-steps/#define-a-path-operation-decorator

# could be that we don't need async eg db call
# https://fastapi.tiangolo.com/async/#in-a-hurry
async def root():
    return {"message": "Hello World"}


# Define a new GET endpoint for /items/foo
@app.get("/items/foo")
async def get_items_foo():
    # Return a JSON response with item information
    return {"item_id": "foo", "item_name": "Foo Item", "description": "This is the foo item"}

alt text

TODO next is see how AI get generate

  • ask AI how to work with AI

  • youtube how to do this thing like paddy

  • get python going on a laptop?
  • sample CRUD endpoints
  • a lightweight ORM (dapper style?) connector to local postgres
  • front end generators like FastAPI-Admin / Piccolo_admin
  • or custom svelte/vue/ raw html/css/js

DEV NOTES BELOW NOT DONE

Next

nginx and use certbot to handle ssl certs

swagger client / tests?

try a gui builder like FastAPI-Admin or piccolo_admin or use svelte

or html/css/js Jinja2

All in one Vibe coding sites

https://bolt.new/~/stackblitz-starters-zotmtg6q

https://v0.dev

synthesia

AI’s

chatgpt

cursor (Claude?)

FOO

for a complex webapp:

v0

shadcn/ui uses radixui tailwindcss

next - lots of pages… (and react). these are the ‘defaults’ of AI

**try different tech stacks with AI

full JS - svelte… next? full server - .NET, Python, RoR, PHP, Rust?

  • Simple
  • No ORMs!
  • Only a few devs, so can be agile and lean

TODO

podcast: https://syntax.fm/show/923/getting-the-most-out-of-ai-coding

linting and formatting - ruff?

Cursor Rules

MVP Single Page Website

single page website saying what we’re doing, to get a launch list of email addresses.

have screenshots of final product (mock ups?)

https://www.figma.com/sites/