Vibe coding apps taught me how hard real coding is

Like so many others, I jumped onto the vibe coding bandwagon, entranced by the idea of building my own incredibly useful apps with nothing but an AI prompt. 

Over the course of about six weeks, I did manage to build my own apps–four of them, to be precise. By my count, three of them pretty much suck.

The one app that does work fairly well–a graphical dashboard that tracks the status of my Docker Swarm setup, a “unified cluster” configuration that allows various software services to dynamically float from computer node to another–took me about a week to vibe code, and the process went fairly smoothly.

My other projects? Not so much. I made three attempts to reverse-engineer OpenClaw, the viral agentic AI sensation that everyone in the AI industry is furiously trying to copy. My purpose was two-fold: one, to build a locked-down version of OpenClaw that I was comfortable installing on my own system, and two, to get some much-needed experience with AI agents in general.

I’ve been working on my “BenClaw” app for weeks, in various forms and using various tools, including Claude Code, Google’s Antigravity, and OpenAI’s Codex. 

Initially I tried adapting existing OpenClaw clones on GitHub, then I tried vibe-coding my own from scratch. “OK Claude” (or Gemini, or Codex), I’d ask, “I want to build my own OpenClaw.” I’d then paint the various features with the broadest of strokes, while giving the AI plenty of latitude in terms of the outcome. (That, as it turned out, was my first of many mistakes.)

I have three vibe-coded prototypes on my test bench right now: an agentic but forgetful AI chatbot that communicates through Discord and frequently complains that it’s exceeded its tool-call count; a GitHub-ready agentic workflow helper that I actually didn’t mean to build and have no clue how to use; and a smolagents-based AI agent that somehow became three separate agents, none of which work reliably.

It was after my latest smolagents debacle that I began asking myself whether I was actually getting anywhere during my hours upon hours of vibe coding, or whether I’d become so dazzled by the sight of Claude Code, Antigravity, and Codex spinning up code that I’d been blinded to the fact I was merely spinning my wheels.

Perhaps my biggest issue was my failure to do what seasoned software developers always do: develop a detailed blueprint of the finished product, complete with features, functionality, its look and feel, and–most importantly–what it’s supposed to do. 

“Specing out” or “scaffolding” software projects is tough work, and while you could certainly hand it over to Claude, Gemini, or ChatGPT, the problem becomes that they start making the critical design decisions, not you, 

The result? My second AI agent project, dubbed “AgentLab” by Claude, which does…well, I’m honestly not sure what it does. You’d have to ask Claude.

What we have here is a classic case of “garbage in, garbage out,” and while tools like Claude Code, Google’s Antigravity, and OpenAI’s Codex can do amazing things when it comes to writing and debugging code, all I’d been doing was handing them garbage–vague prompts with no planning, rigor, or vision. The garbage apps I wound up with shouldn’t have come as a surprise.

Now, this isn’t to say that amateur vibe coders like me should give up and leave it to the experts. Far from it. If you have a vision for a new app and want to craft it yourself, now’s your chance.

Just don’t underestimate the imagination, creativity, and elbow grease required both before and after Claude Code, Antigravity, and Codex work their magic.