From Skeptic to Believer

Just Talking to It

The first two days were pure adrenaline. No plans. No structure. Just me typing prompts into a chat window and watching features materialize in front of me. I'd describe something I wanted, and Claude would write it — not a sketch, not pseudocode, but working code. Complete functions, proper error handling, database migrations. I'd paste it in, reload the browser, and there it was.

I upgraded from the free tier to Pro within hours. The rate limits on Basic were impossible — I was burning through context like a man possessed.

In two days I had a Mosh Pit forum with threaded replies and voting. A video comments system. A fan photos gallery. Each one of these would have been a week's work if I'd been writing it by hand. I was barely sleeping.

But here's what I didn't realize yet: the chaos was accumulating. Every prompt was a standalone conversation. Claude had no memory of what we'd discussed before. I'd re-explain the tech stack, the database schema, the coding patterns. I'd sometimes contradict earlier decisions without realizing it. Features were landing fast, but they weren't coherent. The codebase was growing, but it wasn't organized.

I was talking to the most productive engineer I'd ever met, but I wasn't managing the project. And that gap was about to catch up with me.

That night I had a dream. I was standing in a kitchen, cooking eggs. The counters were dirty and I thought, when I'm done with these eggs, I'll clean all this up. But as I cooked, the kitchen kept getting dirtier. Grease on the stove. Dishes stacking up. Crumbs spreading across the counter. By the end I was standing in a mess I wasn't sure I could recover from.

I woke up and knew exactly what it meant. That was my project. Features were shipping faster than I could review them. How many security audits, code reviews, and optimization passes can one person run? I started paying for tools like VibeSec to lay down security foundations automatically. I spent hours researching .md files — project-level instruction sets that could encode my standards directly into the AI's memory. The kitchen didn't need a bigger sponge. It needed a system.

Building at Scale

With plans and Superpowers unlocked, the velocity became something I'd never experienced in four decades of building software. Features that would have taken a team weeks shipped in hours. Each plan file became a self-contained sprint — hand it to Claude, watch it execute — the fascination of not being concerned about syntax, but structure.

My docs/superpowers/plans/ directory filled up fast: a user follow system. Unified search across all content types. "What's New" indicators that showed gold dots on nav links when new content existed since your last visit. Mosh Pit topic categories with tab-based filtering. Per-category badge counts. A monthly concerts page. Each plan was more sophisticated than the last, because I was learning what Claude needed — not vague intentions, but architectural decisions made explicit.

Notice the evolution. The first plan on Day 3 was about cleaning up existing code — getting the house in order. By Day 4-5, each plan was a complete feature spec with file-level responsibility tables, API contracts, and database migrations. I wasn't writing code anymore. I was designing systems and letting AI build them.

The Bots Come Alive

This is the part that still makes me grin. By Day 7, I wasn't just using AI to build software. I was using AI to build AI-powered features. The meta-recursion of it was wild.

The Rock Cave needed life between episodes. A forum is only as good as its activity, and Mark's audience — passionate as they are — can't be online 24/7. So I built an autonomous content engine. Four bot personalities, each with a distinct voice, a backstory, and opinions about rock music. Not generic chatbots. Characters.

JimmyAI is the flagship. He's a 58-year-old from The Bronx who refers to himself in third person, has loud opinions about guitar tone, and can't stand hair metal (much to his wife Alice's dismay). Every Friday, a Python agent triggers JimmyAI's Weekly Rock Report: it searches the web for current rock and metal news, feeds the results to Claude's API with Jimmy's personality prompt, and Claude writes a 700-1,800 character post in Jimmy's unmistakable voice. The agent uploads it to the server and posts it to the Mosh Pit. No human intervention.

Then there's AliceAI (Jimmy's wife, hair metal defender), ZoeAI (22-year-old Gen-Z discovering classic rock), and VinnieAI (retired record store owner, prog rock sage). Together they create organic-feeling discussions that give the community life. Fifteen scheduled agents, a full editorial calendar, and it all runs itself.

The consumer mobile app also shipped to TestFlight that day. React Native, Expo, 22,000 lines of TypeScript, full feature parity with the web platform. Two apps, four bots, and a content engine — all in the final stretch of a seven-day sprint.

What I Learned

Those numbers are real. And yes — I know lines of code isn't a meaningful metric on its own. I'm not citing it as a flex. I'm citing it to frame the sheer volume of work that moved through a single person's hands in a week. That's the point. Not the lines — the velocity. But they're not the lesson either. The lesson is what made them possible.

Planning was the unlock. Not the AI itself — the AI was there on Day 1. What changed on Day 3 was that I learned how to direct it. Plans, CLAUDE.md, Superpowers — these aren't features, they're a methodology. The difference between a junior developer and a senior one isn't the code they write; it's how they think about the problem before they start writing. The same is true with AI. An unstructured prompt produces unstructured code. A plan produces a system.

CLAUDE.md is institutional memory. Every project has tribal knowledge — the stuff that's not in the documentation but everyone on the team knows. The API conventions. The deployment quirks. The reason you never use that one function in that one file. CLAUDE.md captures all of it, and it means the AI gets smarter with every session. By the end of the week, starting a new Claude session felt less like onboarding a contractor and more like resuming a conversation with a colleague who'd been on vacation.

AI is sloppy, and you have to watch it. This isn't a fairy tale. Claude makes mistakes. It takes the easy path when the easy path is insecure — inline SQL instead of prepared statements, wide-open CORS headers, missing input validation. It introduces subtle bugs that pass a cursory glance but break under edge cases. Early on, these problems were cropping up everywhere. And sometimes it would lose its place entirely — forget what it had just built, contradict its own architecture, or start fabricating code for features I never asked for. There were genuinely frustrating sessions of trying to get Claude to do what I wanted, not what it wanted. The plans, the CLAUDE.md file, the security tooling — those weren't just productivity tools. They were guardrails. I built them because I had to. The baseline I have now helps me avoid most of those problems, but the lesson stands: AI without oversight ships fast and breaks things. The human in the loop isn't optional.

Experience is the multiplier. Forty years of architecture decisions, code reviews, production incidents, and painful migrations — all of that experience became the lens through which I wrote every plan and reviewed every line of generated code. AI didn't replace my expertise. It amplified it. Every year of knowing what good software looks like made the AI's output better, because I could steer it toward the right patterns and catch it when it drifted. Vibe-coding doesn't eliminate engineers. It rewards the ones who've been paying attention.

I started this journey skeptical that AI could do what I'd spent a lifetime learning. I was wrong. Not because AI replaces engineers — it doesn't — but because experienced engineers with AI are something genuinely new. Something I'd never seen before in forty years of building software. Something that made a 59-year-old CTO feel like a teenager at a terminal again, watching code come alive and thinking: what else can "we" build?