Reflections & Musings On The First AI-Assisted Engineering Project

In July 2025, I was trying to figure out how to distribute the documentary for my debut album, Dream River. I had all these stories and a 45 minute documentary about the journey of making this album. Just throwing it on YouTube felt wrong. So I started digging.

I knew YouTube had an API that emitted timestamps for chapter markers. I wanted to see if it let me use it and also if I could control the content reliably through it on my own web page if this was embedded. After doing a bunch of research and a few JSON files later, the shape of my idea was definitely possible. Have the documentary play on the webpage and have dynamic content under the documentary that gave you an even deeper insight and commentary through the page.

The problem was I had not written a single line of HTML since 2014, the idea of using Stack Overflow now 11 years later seemed crazy, and every website building service was a drag and drop editor inaccessible to me as a blind user. But this whole AI thing was gaining traction. Every company was like guys, AI can code now. Mind you this is July 2025, Opus 4.5 wasn’t out yet and agents were a big question mark.

But I had a $20 ChatGPT sub and I vomited a prompt out and 10 or so minutes later it gave me a single HTML file with everything nested inside. It was really really messy. I tried reading through it and gave up after 20 lines. I opened it though, and the prototype was actually fine. This was GPT 4.x, I don’t really remember the exact model. So the gears started churning, maybe I didn’t need Stack Overflow, I could copy paste from GPT and get to the end.

However, the thought of manually copy pasting hundreds of blocks of code, arranging them in the right file structure and testing things immediately made me go, well you know what, let me just play music. That same week though, I downloaded Cursor.

And that was like magic. This was before Cursor got as bloated as it is today. The accessibility was actually navigable back then in Cursor and their Electron didn’t kill me. Cursor was my first agentic experience. What we now take for granted, back in July was like magic. This thing would actually just run on my system if I let it. I wasn’t brave enough for YOLO mode, but damn was it great. Only for a bit.

While Cursor had agentic capabilities, the models were still what they were in July and August of 2025. This single page experience took me 3 weeks to make. The models didn’t have the ability to self verify. The vision models weren’t good enough to accurately get stuff from screenshots and most of all I only understood basic HTML and all this JavaScript it was writing was out of my wheelhouse.

The Cursor window on one side, the ChatGPT website on one side and the constant copy pasting of JavaScript to have it be explained to me is a time I do not miss at all.

The page was up in three weeks though and I went from not knowing how to do this, to being able to implement it and deploy it as a live website in 3 weeks. This was further than I got when I studied HTML and web design in college over a whole semester.

Most importantly though, and I am glad for this experience, it taught me really early on that you HAVE to verify AI code. There is a debate going around now in 2026 whether you should read the code or not. Some people I really admire, like Mitchell Hashimoto, read the code. Some people I think are trailblazing the space, like Lauren Tan, have PRs merge when they sleep.

I honestly don’t care for this debate because whatever method of verifying the code works for you is the one you should choose, but both of these go back to my early take away: you HAVE to verify the code.

I couldn’t read all the syntax accurately, so my initial method of verification was to paste the code into a fresh chat window and make sure it was doing what I asked of it, which is now called an adversarial agentic review with fresh context. The Gemini models had decent vision, which I was using as a proxy for my own eyes since I got none, and this today models can just do. And I was using scripts to automate the accessibility workflow till it got it right, which now is called a loop.

All in all, this webpage still exists in the shape I made it back then. It’s rough around the edges, the mobile version is janky at best, the CSS isn’t the most elegant, and it’s disjointed from the rest of the website. I have purposely not gone and revised this even though it will probably take an hour or two today. It’s an artefact for me to look back on how things started.

It’s nice to reflect on what has stayed the same and what has changed. In a lot of ways, everything is the same. I still don’t flawlessly read any syntax and get agents to explain things to me. I still use vision models as proxies for my own eyes and I still run verification loops. Inside those steps though everything has changed. I’ve read a ton of books and done courses to understand architecture, my vision models get a checklist and a rating list that they have to fill out and send back, and the loops have a lot more computer use and linters built in to make sure the code is verified.

But most of all I’ve built a lot over the last year and there is just no substitute for getting your hands dirty. It’s interesting to me as a person coming in from the outside how things look and seeing both sides of the argument and belonging to neither. I don’t know what the future holds for people like me who always wanted to learn something but didn’t have access to the resources to learn that thing the way they wanted.

What is the combination of having things taught to you the way you want plus having the tools to execute that are malleable to the way you work? While I do think software is in a unique place for AI to work extremely well, especially because of the verification loop, it also terrifies me when people extrapolate that to other fields with no verification loop. I have asked some not so basic music theory questions to every AI model for example and they would fail my undergrad music theory exams. So while I’m optimistic about the future, and I want to believe that the world is eventually going to be a better place with this technology in it, I really hope fellow people take Uncle Ben’s words to heart: “With great power comes great responsibility.”

Wait, wasn’t I talking about building a website? Oh well. Until next time.

Eshaan.

P.S. If someone can actually get to the people at Cursor, guys I would love to make your app accessible again. I really actually miss using Cursor.