Fast & Furious Code With AI
AI at the Wheel, Craft in the Rearview?
Following is an AI generated blog from a transcript of a lightning talk I did recently at Incubyte.
Over the past few days, I embarked on a coding adventure using AI, and it blew me away. A year ago, I tried AI for development, and it was a little like giving your toddler the keys to the car—lots of mistakes.
Fast forward to today, and AI is like a highly caffeinated intern that codes faster than I can keep up. 90% of my latest project was handled by AI, and what used to take weeks was done in four days—even while juggling meetings!
Here’s how it all went down.
My Journey with AI: From Concept to Launch
Let me introduce you to CodeAid MVP Copilot. This little app is a product creation assistant for founders that helps you validate your startup ideas, conduct market research, and even helps create product documentation—press releases included (because, you know, working backwards from Amazon’s dream day is apparently a thing now).
And guess what? AI did most of the work. Seriously, 80-90% of this was AI-driven. I only had to step in to adjust prompts and check results. I could literally tell it to create a wireframe or "imagine it’s a world-class UI designer" and it would generate layouts that matched the exact vision I had. It's like having an intern who’s scarily good at taking orders.
You can see the working product at CodeAid MVP Copilot.
Top 3 Advantages of AI
Speed: Seriously, it’s like strapping a rocket to your code.
Speed: Did I mention speed?
More speed: AI doesn’t need coffee breaks or sanity checks—it just keeps going.
Why is this game-changing?
What took me four or five days with AI might have taken weeks without it. And the best part? I could focus on refining the finer details that usually get left for later—like auto-scroll or rounding UI corners. Normally, these are the details you never get to until your MVP is functionally ready!
Things I Noticed Along the Way
Anthropic’s AI: If you haven’t tried it, you’re missing out. Anthropic is the new player, and it gives cleaner, more accurate results than some of the others I tried (looking at you, Gemini).
LLM understanding complex programming principles: I discovered that the AI wasn’t just coding blindly—it actually understands complex programming concepts like DDD, SOLID principles, and Clean Code. Add what you need but also how you need it, and AI will grant your wish.
Prompts are everything: This can’t be overstated. The quality of your results is directly tied to the quality of your prompts. Vague inputs lead to vague outputs. I found myself doing iterative prompts—gradually refining them until I got exactly what I wanted. Sometimes, it was as simple as adding a line like, "Use SOLID principles" or "create separate React components" to get better code. Eventually, I was writing prompts that were several lines long, almost like mini-documents.
Known tech vs. cool tech: I tried using frameworks like Bulma, and it just didn’t play nice. But as soon as I switched to Tailwind, it was like the AI was speaking its native language. Stick with the frameworks AI has "seen" a lot, and it’ll save you a ton of time.
LLM creating prompts for itself: One of the more mind-bending things I discovered was how it could actually create prompts for itself. At one point, I asked the LLM to write a prompt for me to generate specific code, and it worked like a charm. This feature takes AI assistance to another level—it’s like having a co-pilot who not only takes instructions but also suggests the best way to ask for those instructions.
Challenges & Concerns: Craft Takes a Backseat
Craftsmanship: AI made me lazy. The craft of coding took a backseat as I let AI churn out code. While it generated functional code, I wasn’t as involved in the details, especially with test cases. The AI generated test cases for me, but I skipped my usual test-driven development (TDD) practice. This made me nervous because I couldn’t be 100% sure all behavior was covered under the tests.
AI creating integration tests: The AI did create functional integration tests, but they were basic. These tests will work, but they’ll likely need refinement, especially when handling more complex scenarios. The speed was great, but the human touch was missed here.
Using AI to fake integrations: Another cool thing I tried was having the AI generate fake integrations. This acted as a placeholder for real integrations while I developed other features. It’s a handy trick when you need something to hold the place without fully implementing it right away.
Privacy and PII concerns: AI is powerful, but enterprise readiness is still a concern. When using AI, especially with sensitive data, we have to be mindful of privacy and PII issues. It’s why we’re considering using our own hosted LLM at Incubyte to make sure data never leaves our control.
Short context vs. long context with AI
Another interesting discovery was the balance between short context vs. long context. Feeding too much data to the AI all at once—like keeping a long chat history or overwhelming it with files—can actually hurt the quality of the results. I found it much more effective to break the tasks into smaller, focused conversations, giving the AI clear and concise instructions for each step. It’s a great reminder that, while AI is powerful, it still needs us to structure its workload thoughtfully.
Wrapping Up: AI + Craft = The Sweet Spot
AI is fast, powerful, and increasingly reliable. But the craft of coding—the attention to detail, the nuanced decisions—that’s still something AI can’t fully replicate. At Incubyte, we’re learning to balance both worlds. AI gives us the speed and efficiency, while we keep our eyes on quality and craftsmanship.
So, is AI the future of development? Absolutely. But it’s still up to us to make sure the wheels don’t fall off along the way.