From Legacy Code to AI-Powered Development

At my previous job, I was a lead developer on a product that generated 40% of the company's revenue. With the help of a coworker who had served in the same role for a couple of years, I mastered an undocumented legacy codebase, squashed countless bugs in spaghetti code, conducted major migrations, and built new systems that interfaced with other apps.

I also gained experience with modern tooling when we transitioned from Vue 2 to Vue 3, using Storybook (which, honestly, I found to be a distraction), Vite, and other contemporary JavaScript tools.

My Journey with AI Assistance

I began using AI for coding because of those moments of magic where I thought, "Wow, this is really helpful!" There were delightful instances where previously laborious tasks became mindless and significantly faster.

I started with GitHub Copilot but was frustrated by its limited context and inability to handle more than 200 lines of code. I even used AI in the terminal with fig.io (now owned by Amazon), which made command-line work amazingly efficient. I'm looking forward to trying Warp terminal's AI features next.

For larger coding challenges, I used the web interfaces of Claude and ChatGPT, pasting in code snippets. Claude was particularly cool because it often attached code as a document, creating a better UX. While ChatGPT was solid, I kept hearing Claude 3.5 was superior, so I began using it extensively.

Claude 3.5 Sonnet has been consistently great for me. Claude 3.7 can be really good sometimes, but other times it takes too long and gets stuck in loops. I should try other models to understand their limitations and strengths, but that takes time and experience.

Enter Cursor AI and Vibe Coding

When I discovered Cursor AI, essentially a VS Code fork with vastly improved AI capabilities, everything changed. "Vibe coding," a term coined by Andrej Karpathy, describes this new workflow where you express your intentions to AI and then refine what it generates.

My productivity skyrocketed—closer to 100x than 10x. What would have taken weeks was accomplished in a day. My portfolio now features:

  • Responsive design
  • Light and dark mode
  • Smooth animations and hover effects
  • Clean, hand-crafted code

Tips and Tricks for Vibe Coding

1. Simplify Your Tech Stack

Ironically, I found the most success when I abandoned complex frameworks and returned to vanilla HTML, CSS, and JavaScript. The AI struggled with React, Next.js, Vue, and Nuxt projects, but excelled with the fundamentals.

2. Create a .cursorrules File

Place this in your project root to guide the AI. I prompt engineered mine to emphasize clean code, DRY principles, and efficiency. This drastically improved the quality of generated code.

A great resource is Cursor.directory, which offers various .cursorrules files for TypeScript projects, Frontend development, Next.js, React, and more. I'll feature more about their MCP servers in my next blog post.

3. Choose the Right Model

I typically use Claude 3.5 Sonnet for most tasks - it's efficient and gets more done. Occasionally I'll switch to 3.7 when I need more reasoning capabilities, but 3.5 seems to hit the sweet spot for coding tasks.

4. Use the "Restore Checkpoint" Feature

A technique I've learned is that you can hit the "restore checkpoint" button under older prompts to revert to a previous state. This form of git/version control is a great tool and saves a lot of time. You can even edit your instructions to avoid subsequent errors—an advanced technique that Cursor's UI allows.

5. Become a Codepreneur, Not a Code Monkey

With vibe coding, you don't review every line of code anymore. Instead, you selectively test like a playwright, checking the browser and web app for interaction errors and visual bugs.

You're now conducting an indefatigable, autistic genius who's brilliant in some ways and totally clueless in others. You're not a code monkey anymore—the AI is your code monkey. You've become a codepreneur, a technical director, product manager, and more.

Speech-to-Text: The Perfect Companion

Another game-changer in my workflow is speech-to-text software. I was using Whispr Flow until my trial expired, and now I'm using Mac's built-in speech-to-text tool. It's decent, but definitely not as good as Whispr Flow (though I'm not sure it's worth the price tag yet).

The beauty of speech-to-text is that you're not limited by typing speed and you don't experience typing fatigue. When combined with vibe coding, it creates an incredibly fluid experience—you speak your intentions, AI implements them, and you direct the process.

What's Next: MCP Servers and Ambitious Projects

After this blog post, my next task is to write about MCP (Multi-Context Programming) servers—a technology that allows LLMs to reach out to external data sources, APIs, and services in real-time. These servers extend AI capabilities by enabling them to perform actions outside their training data, maintain context across multiple interactions, and support tool use. In Cursor specifically, they can help access documentation, search package repositories, and maintain programming context across multiple files. It's a very new development, and in this era, you've really got to keep your ear to the ground for these big movements.

Under the "Selected Work" section of my portfolio, I have three project cards that need filling out. I'm planning to ask AI directly: "I want to use Supabase for my backend—what would be a good project idea?" For another card, I might even try Unity and see what the AI suggests I create.

That's the beauty of vibe coding—you can be as ambitious as you want since mistakes cost so little time. You might as well try something bold! I'm definitely going to produce some super cool full-stack products, MVPs, and prototypes based on ideas that I'll work with AI to refine and improve.

The Bottom Line

There's nothing more I can say than this: these are exciting times. Vibe coding IS an essential part of coding as far as I'm concerned. In just a day, I've pumped out a pretty nice developer portfolio, and it's only going to get better. This will be my second published blog post on the site.

The AI revolution in coding is here, and if you embrace it, you might find yourself rapidly creating projects you never thought possible—while directing the show rather than typing every line yourself.