I Vibecoded My Way Through My Internship…
Here's some lessons I learned this summer from building work projects using AI tools
Over the past summer, I worked at a venture capital firm to investigate how to use AI in their work. I built anything from Salesforce automations to log updates on companies to an intelligent search engine for deal sourcing. As someone who didn’t decide to take up a technical major until junior year, it was a dream to be in an investing and engineering role. That being said, it was an uphill battle:
No one else in the firm had an engineering background, so no one could evaluate my code
Taking 10+ demos from existing vendors made me realize one-size-fits-all solutions weren’t a good fit, and that I’d have to develop new software
I’d been following the progression of AI-coding assistants like Cursor for months and saw the true power of generating hundreds of lines of code within seconds when I participated in a hackathon during the summer.
So I opened my AI coding assistant and started vibecoding…and it was kinda terrible. Several lines of sloppy code would be generated for one feature that could’ve been implemented in a couple of lines. Cursor would glaze me half the time without actually addressing a bug I found. I’d spend days figuring out how to make a successful API call.
But eventually I figured it out, or at least ended up learning to use these tools more effectively. Here are some takeaways below.
Best Practices
1️⃣ Choosing the Right Tech Stack for the Job
Not every project needs the latest framework. Some needed quick-and-dirty speed; others had to be production-ready. I picked stacks for integration ease, maintainability, and cost effectiveness.
2️⃣ Prompt Engineering as a Skill
Good prompts are like good requirements for a teammate — except this teammate sometimes hallucinates. I learned to start small, test quickly, then add constraints. I suggest reviewing this video on how to optimize your prompts.
3️⃣ Finding the Right Existing Tools
Sometimes the smartest move was finding an API, repo, or feature that already existed. Vibecoding is knowing when not to reinvent the wheel.
4️⃣ Understanding the Customer
The tools I built weren’t for engineers — they were for investors and portfolio managers who needed insights, not code. If it didn’t solve their problem, it didn’t matter how elegant the code was. This mindset applies to communicating your work as well - I’d always share the insights and results of my work first, then the process, because the former was more relevant to the investors at the firm.
Mysteries & Roadblocks
Even with successes in building out my internship projects, I kept running into unsolved puzzles:
QA for AI-Generated Code – “vibe-debugging” doesn’t exist yet. Even with AI, I still had to know Python, JavaScript, data structures, and basic AI/ML just to be efficient with debugging code. If we could do this entirely via natural language, this would further lower the barrier to programming.
Auditable AI – Observability platforms exist, but LLMs that serve as the foundation of these AI projects remain “black boxes” where you can’t see exactly how they reason through problems.
Memory – What’s the best way to store it? How do you integrate it into prompts without drowning in noise?
AI “Industry Expert” Intelligence – Models can fake confidence, but they can’t be the expert yet. I had many instances where GPT or Claude would give me a whole story behind a startup or industry it researched, just for it to be mostly hallucinated, inaccurate text.
The Bottom Line
Vibecoding isn’t just sending half-baked prompts.
You have to be intentional about when to lean on AI for momentum and when to step in and steer hard. And while AI can get you 80% of the way, it won’t get you to the other 20% of building and shipping software, on top of the planning that goes into system design or UI/UX.
But the amazing thing about vibecoding is that it’s especially easy to get started! You don’t need a perfect project or a job to start building with AI.
Last weekend I threw together a birthday wishlist tracker and a hydration reminder bot within a couple of hours! Imagine all the mini-apps you could build over a weekend.
Try building something with AI tools; all it takes to start is a few sentences on a chatbot.