Confessions of an Accidental AI Developer

Hi, and welcome to The Atomic Builder! 👋

Last issue, I introduced Feedback Force, the AI-powered feedback analysis tool I self-built, using a force-directed graph to tell a story about user feedback.

Today, I want to take you behind the scenes. This is the messy, funny, and slightly terrifying journey of building it. Think Netflix’s "The Making Of" - except I’m not getting royalties…yet.

Netflix, call me…?

Stepping into the Lion's Den (Hello, Cursor!) 🦁

Let's get this straight - I’m not a developer, not even close. I am however an adventurer, with an endless appetite for chaos. I decided to build Feedback Force using Cursor, a developer's best friend, to give me more control. On their website, they describe Cursor as ‘the best way to code with AI’ - so it’s blatantly not aimed at non coders, but I wanted to see how far I could get…

Was the complexity worth it and should you consider it?

Let's dive in.

🛠 The Build: Real-Life Debugging Drama

Initial Setup:

• Before starting in Cursor, I asked Claude to guide me. Claude gave me the instructions for Cursor to create my app - that was a crucial starting point - as it gave it the structure my app needed.

• Once I had this starting prompt, in Cursor, I slalomed through npm commands, endlessly typing ‘npm install’, desperately hoping something would click. Spoiler alert: it didn’t. Again, Claude came to the rescue.

• "Package not found" errors popped up left, right and centre - Cursor’s way of saying, “Install xyz” & “Help me, help you!”

• For extra help, I used the chat feature in Cursor, that I tag teamed with Claude, to get me out of the mess (of my own making) I often found myself in.

User Interface Adjustments:

• My first version of the app in Cursor was basic and (somewhat) functional, you can see it below in my local environment.

• The feedback mini-map had a mind of its own, stubbornly overlapping the graph, then placing randomly out of screen.

• I needed help. I paired Cursor with Grok 3 (xAI’s chatbot) and ChatGPT (my external counsel!) and it turned out to be an excellent decision - buddying up AI tools for troubleshooting is a powerful hack if your AI generated code is failing.

• I thought a "weight slider" would be a great idea—until adjusting it became a full-time job. I couldn’t get it working so I ditched it.

AI Integration (When Things Got Interesting):

• Importing small datasets to test the app worked like a charm. Larger datasets sent the app spiraling into existential panic mode - something still on my fix list.

• My integration of AI sentiment analysis wasn’t as magical as hoped - more like a stubborn intern that needed precise instructions. "Make it work better" wasn’t cutting it. The AI needed structured prompts, or it just did its own thing.

Visualizing Feedback:

• The concept I had of a force-directed graph sounded amazing - the reality of getting it working well, another story, the nodes were, what can only be described as, convulsing when clicked... Still figuring that one out.

• The dashboards help capture insights that the graph can’t, but scaling up to handle large datasets remains a challenge.

Some early screenshots of the app

I tested a first version of the app early on with limited data and functionality, and also a different name, which I didn’t like so I ultimately switched out ‘User Research Insights Network’ for ‘Feedback Force’!

🎯 What I Did Well

As you can see, the journey was wild. Ultimately I finished with something somewhat in the realm of MVP. I think you can afford me that, right? This is what I think went well:

1. Stepping Boldly into Complexity:

   I embraced complexity for greater control, challenging myself beyond simpler no-code tools to deeply understand the MVP creation process and some development principles - logic. This was priceless in helping me learn.

2. Systematic Debugging:

   I patiently tackled UI, dependency issues, and AI integration, step by step, methodically learning from every (excruciatingly!) frustrating moment. Seeking help from other tools in combination with Cursor unblocked many obstacles.

3. Interactive Feedback Insights:

   I successfully created dynamic interactions, enabling actionable insights directly from visual nodes - no Excel (shudder…) sheets harmed in the process.

🚨 Mistakes I Made (So You Don’t Have To)

The more I build with these tools, The more I learn about the nuances and tweaks that help get you where you need to be, faster. These are 3 things I’ll think through more before my next build:

1. Underestimating Technical Complexity:

   I assumed Cursor would simplify AI coding. My reality check: It introduced dependencies and cryptic errors that slowed me down significantly.

2. AI Integration Without Clear Rules:

   I didn't initially structure clear (enough) prompts for AI sentiment analysis, leading to confusion, extra rounds of fixes, and unnecessary rework.

3. Feature Overload (Goodbye K.I.S.S!):

   Early on, I added way too many complex controls (like "Pause Simulation" and "Connect Nodes"). It took a while to untangle, but I settled on simplicity. It wins every time.

Snippet Of My Chats with Cursor

The early days of coding with AI are filled with moments that you frankly couldn’t make up. I wanted to share some of the snippets here of my chats with Cursor…

If nothing else, I hope it will help you understand where we are and give you confidence to venture forward, no matter how challenging you may find it, today.

🚧 What's Next? (What I'd Still Love to Fix):

• Tree Shaking & Optimization: Node shaking still feels shakier than my confidence when deploying code.

• Landing Page Demo: A polished demo directly on the landing page would be awesome.

• Scaling for Big Data: I built for about 100 comments. But what about 1,000? Challenge not yet accepted.

Will I? Probably not at this time...

Now, Your Turn

Actually, I’m not sure that asking you to dive into Cursor, yourself, without any training wheels is really good idea.

In fact, I don’t advocate that you do that at all!

Final Thoughts

This experience proves non-technical creators can leverage AI to build amazing things, even at the more challenging end of the spectrum - though it’s genuinely VERY challenging. Right now, it's more brute force with tools like Cursor (to be fair I am not its intended audience).

The real power for you, comes from creatively combining the right tools to build things. Whether you're crafting ambitious apps, quick prototypes, or just asking ChatGPT to diplomatically email your kid’s teacher (no judgment here), remember: the tools are there. Combine them to your advantage. That’s where I (think!) I excelled here.

Now, I better check my phone, I’m waiting on a call from Netflix…

Until next time, keep experimenting, keep building, and as always - stay atomic. 👊

Faisal

This Week’s Build Beats 🎵

Each issue, we pair the newsletter with a track to keep you inspired while you build.

This week, because AI building sometimes feels like you’re halfway there, especially when typing npm install for the hundredth time…

🎧 ‘‘Livin’ on a Prayer” – Bon Jovi

Grab the playlist on Spotify - I add to it each week!