My Experiment With Vibe Coding

Last night I built a full-stack web app in under ten hours. I’m not a software engineer and have near zero development experience, yet I shipped a tool that collects meeting inputs through a simple form, pulls details from internal directory and calendar APIs, formats a brief, and emails it to the intended recipients. I got the job done in a single extended evening using Cline inside VS Code for orchestration, Harmony for the app framework, and JDK 17 for the runtime. I also used MCP, the Model Context Protocol that lets an AI agent securely connect to tools, files, and services without hard-coding every integration.

Before this, I had used AI to build and refine this website: Hugo for the static site, VS Code for writing, v0 by Vercel for quick UI sketches, GitHub for versioning, and Netlify for deploys. As a product manager, I have also been using Claude in Amazon Bedrock to tighten specs and draft plain-language strategy artifacts. For analytics, I anchor data in Amazon Redshift and ask the model to propose cleaner SQL first, which removes a lot of tedious effort such as creating joins, CASE statements, window-functions, and the like. This process, describe intent cleanly then implement, translated directly to my work on build the app.

I ran the work as a string of short modular sessions in Cline, each with one goal and a set of simple success checks. We started with semantic HTML and core flows so the app stayed accessible and failed gracefully, then layered JavaScript where it was essential e.g., search bars and the like. We initially used mock services for directory and calendar and swapped in the real APIs via environment toggles when access issues were resolved. I integrated one API per session, added error handling, fallbacks, and clear user messages, and ended every session in a working state with validation notes and updated docs so the next step started clean.

Cline got me through the hardest stretches with ease. We pivoted mid-build to hosting the app on Harmony, and Cline swapped mock calls for production one at a time. It also scaffolded the data layer and validation, wired dependencies, and kept docs and build configs in sync as I optimized the core logic at a furious pace. We tuned live employee search with a 300 ms debounce, a minimum-character rule, and clear loading states. When we hit permission and security clearance walls at 95% readiness, I felt the pain my software engineering colleagues often talk about but one that I had never experienced first hand before this. The end result was a working prototype with backend logic, basic auth, a real database layer, and live docs built by a product manager who didn’t really code until now.

This experience changed how I think about the “AI bubble.” One recent work by MIT says many AI pilots fail because tools don’t learn or integrate. See coverage of MIT’s State of AI in Business 2025 report, for example in Forbes. Another says rule-based work is about to shift faster than incumbents expect. See Jonathan Gray in the Financial Times. My night-build is a small, hands-on proof of the latter and an answer to the former: a non-developer shipped a useful tool by keeping steps small and context shared. More importantly, it felt natural to someone trained to think in systems. Years of writing specifications had already taught me to describe behavior step by step, and the AI agents made those instructions executable. My value moved from interpreting requirements to designing precise instructions that compile into working software.

Today, product manager jobs are rapidly shifting from authoring specs and strategy papers to shaping live, testable systems. The technical bar to create is lower, while the cognitive bar i.e., clarity of logic and precision of expression, is higher. Hype, failure, and wasted capital will likely still happen, as with every tech wave that births a new industry. What’s different now is the immediacy of utility even in the infancy stage of the technology. When someone like me with no coding background can build a full-stack application in less than a day, the “AI is a speculative bubble” story stops matching reality. It feels less like hype and more like the arrival of disruptive infrastructure. And while I make no claims about the prospects of AI super intelligence or its alarmingly high natural resource costs, I think the productivity enhancements from AI tooling available and in the works today, if sustained, will multiply human capability manifold in the next 3 to 5 years.