Just How Close are AI Coding Assistants to Helping Non-Coders Ship Products?

Summary: AI coding assistants have arrived for front end prototyping/ landing page building/ mini apps but anything requiring persistence or server side work is not quite ready for no code builders like me.

Background: As a no code builder with over 10 years experience building with the help of others I've been eagerly awaiting the promised AI coder tools. I've been testing out every tool that promises to help me code so my takeaways are below. So you can baseline against your skill level, I can't code but am familiar with no code tools and am comfortable building workflows with Zapier. I've been working with Engineers for 15+ years so know high level what is possible but have no ability to do anything server side.

AI Coding Tool Overview: While there are many tools popping up daily, the landscape as I see it breaks down to the following tools and capabilities (note my personal reviews below are marked with an * on the list below):

1. AI coding assistants for IDEs (designed to make professional developers more efficient)

   1. Cursor

   2. Github Copilot

   3. Supercoder

2. Coding Assistants for Front end work (landing pages, UX, simple prototyping)

   1. Screenshot to code

   2. TLDRaw Makereal*

   3. LLM + prompting + Pick your IDE

      1. Gemini

      2. GPT4o

      3. Claude Sonnet 3.5

   4. Vercel V0

3. Natural Language Chat with AI Code Previews

   1. Claude3.5 (w/ artifacts)*

   2. Poe Previews*

   3. GPTEngineer*

   4. Pythagora (vscode plugin: yes. link4 )

   5. https://www.blackbox.ai/ Robocoder (vscode plugin: yes link)

4. Full stack w/ own UX/ IDEA

   1. Replit*
      

Takeaways

• The LLMs with contained code execution window/ preview windows are the best becauser they allow quick iteration especially on the front end interaction

• Knowing how to write user stories like a PM is really really helpful as thorough descriptions of what you want to do (whether it be interactions, flows, design) increases the likelihood of success dramatically

• Haven't gotten access to Devin yet which I hear is the best

• Right now the server side is not solved. Could set up client side tools but these are not production grade as they are slow, don't persist, and get fucked up if the user closes the session

• Using LLMs for code outputs is very tedious and requires forcing the outputs to be full code snippets. They also lose context at times so require pasting full code snippets in

• Most of the LLMs break down at some point and can't solve bugs and get into a loop

• Live chat+preview UX feels the most intuitive and indeed the future of where these tools are going for non coders

My notes on using each of the AI Coding Tools:

Replit - while Replit had a few major benefits like 1) forking existing working projects, 2) help with package installation, and 3) the code preview tools I found I had a hard time debugging issues once I diverged too far from the original forked projects. Here it became hard to tease out if the bugs were on the Replit side or my own (especially when using LLMs like Gemini or GPT4o). I just found that the added question of "is the bug on Replit's side" was not worth the positives above.

GPTEngineer - Highlights for me include 1) a chat + almost real time preview, 2) uses Github code versioning which gives a nice bit of visibility into what is happening which helps when an AI is coding for you. It does requires you to set up Github and projects are Public and shared to a feed by default which I don't love when fumbling around

Claude Sonnet 3.5 w/ Artifacts

Poe Previews

TLDRAW MakeReal