Vibe Engineering {VE} #6 — The Experiment
The real story of how a young SRE, a wild idea, a few fragile tools, and a lot of persistence led to something we never expected.
🔥 This post is different from the others.
No frameworks. No analogies. No big philosophical statements.
Just the real story — exactly as it happened.
The experiments. The failures. The moments of doubt. The breakthroughs. The people who thought I was crazy. And the moment when something we built actually went beyond what we had imagined.
This is the unfiltered origin of Vibe Engineering.
⚡ It started with Karthik
Late 2022. ChatGPT had just arrived.
Everyone was using it to ask questions. To summarise documents. To write emails. The usual.
I was doing something different.
I had a young SRE on my team named Karthik. Sharp. Curious. The kind of person who asks questions nobody else thinks to ask. And most importantly — willing to go on a crazy journey with a leader who had a reputation for wild ideas.
I pulled him aside one day and said something that probably sounded completely absurd at the time.
"Karthik — I don't think we should just use AI to generate code. I think we should go further. AI should generate the code, execute it, and then improve it based on our feedback. A continuous loop. Human-guided. Machine-powered."
He looked at me the way people look at someone who has just suggested something that is either brilliant or completely unhinged.
😄 He had seen that look from me before.
🔄 The idea that was too early
What I was describing in late 2022 was something that barely existed as a concept — let alone as a working tool.
I was talking about a system where:
AI generates code based on a human-defined intent. That code gets executed automatically. The output gets evaluated — by humans, by tests, by real feedback. And that feedback loops back into the AI — improving the next generation.
Reinforcement learning from human feedback (RLHF). Applied not to training a model — but to our daily engineering work.
At the time, the tools to do this properly didn't exist.
We knew it. And we tried anyway.
🛠️ The fragile early experiments
Karthik and I started experimenting with whatever was available.
We tried several open-source tools. Some were promising on paper. Most were fragile in practice. They would work beautifully in a demo and break completely in a real workflow. The feedback loops we were imagining were still years away from being practical.
Noordeen and Aravindan — two other team members who had seen enough of my crazy ideas to know when to be skeptical — watched these experiments with a mix of curiosity and quiet doubt.
😄 They weren't wrong to be skeptical. Most of the early experiments failed.
But here is the thing about failure in experimentation.
Every failure told us something. Every broken tool showed us what the right tool needed to do. Every dead end refined the direction.
We weren't failing. We were mapping the territory.
🔥 The Aider breakthrough
Then we found Aider.
Aider is an AI pair programming tool that works directly in your terminal — connecting to your codebase, understanding context, and making changes based on natural language instructions.
The first time we used it properly — something shifted.
This wasn't fragile. This wasn't a demo that broke in real conditions. This was practical. This was usable. This was a tool that actually fit the workflow we had been trying to build for months.
We started small. Simple tasks. Getting comfortable with how it thought, how it responded, how it handled context and ambiguity.
Then we pushed it.
We gave it complex problems. Multi-file changes. Refactoring tasks that would have taken a developer days. Bug investigations that required understanding sprawling codebases.
💡 And it did them. Not perfectly. Not without guidance. But it did them — in a fraction of the time, with a fraction of the effort, and with a kind of systematic thoroughness that even experienced developers struggle to maintain consistently.
The loop we had been imagining was starting to work.
Human defines the intent. Aider executes. Human reviews and guides. Aider iterates. The output gets better with every cycle.
This was HumAI in practice. This was D3O in motion. This was Vibe Engineering — not as a concept, but as something we were actually living.
🧠 What we learned that nobody talks about
Through all of these experiments — the failures, the breakthroughs, the late conversations with Karthik about where this was all going — we learned things that I have never seen written down anywhere.
Lesson 1 — Clarity is the most important input.
The quality of what AI produces is almost entirely determined by the quality of what you put in. Vague instructions produce vague results. Clear, specific, well-designed prompts produce something you can actually use. Design first. Always.
Lesson 2 — The human review loop is non-negotiable.
The engineers who struggled with AI were the ones who generated code and shipped it without reviewing it properly. The ones who thrived treated every AI output as a first draft — something to evaluate, refine, and improve with human judgment before it became real.
Lesson 3 — Iteration beats perfection every time.
The worst thing you can do with AI is try to get the perfect output in one shot. The best thing you can do is get a good-enough output quickly — and then iterate. The loop is the point. Not the first result.
Lesson 4 — Skeptics become the best practitioners.
Noordeen and Aravindan — the ones who watched our early experiments with quiet doubt — eventually became some of the most effective Vibe Engineering practitioners on the team. Because they came to it with critical eyes. They didn't accept AI output blindly. They questioned it, tested it, pushed it. And that made them better at it than anyone who had just been excited from the beginning.
🚀 Beyond what we imagined
Here is the part of the story that still surprises me.
When Karthik and I started those early experiments in late 2022 — I had a vision of what we were trying to build. A loop. A system. A way of working with AI that went beyond just generating code.
What we eventually arrived at went beyond even that vision.
Not just in capability — but in what it meant for who could participate.
The insights from those experiments didn't just make our engineering team faster. They started changing how non-engineers on the team thought about building. DevOps engineers who had never considered themselves application developers started building tools. QA engineers started automating things that had been manual for years.
The experiment that started with one leader and one curious SRE — became something that touched the entire team.
And that — more than any specific tool or technique — is what convinced me that Vibe Engineering was real.
Not because the technology was impressive.
But because it changed what people believed was possible for them.
🛡️ That's the only thing that has ever mattered to Fury. Not the tech. Not the tools. Whether the people around him believe in what they're capable of.
🛡️ A word about how this series is written
You just read the real story — messy, honest, unfiltered.
This post was written exactly the same way the experiments happened. Raw thinking first. Structure and clarity shaped by AI after.
This series is openly co-authored with AI — and I want to be specific about how, because the how matters.
Think about the greatest biographies you've ever read. The ones that felt alive, vivid, deeply human. In most cases, those books weren't written entirely by the person whose name is on the cover.
The thoughts, the experiences, the voice — entirely theirs. The craft of translating that voice onto the page — shaped by a writer who worked closely alongside them.
That writer rarely got credited. 🤐
I'm doing this differently.
🎙️ ChatGPT Voice captures my raw thinking as I speak — unfiltered, unscripted, exactly as it comes out of my head in the moment.
✍️ Claude AI then takes that transcript and acts as writer and editor — structuring the ideas, sharpening the language, and making sure what you read actually reflects what I meant to say.
🎨 Google Gemini Nano Banana Pro crafts the cover image for each post — bringing the visual identity of Vibe Engineering to life.
The thinking is mine. The voice is mine. The 27 years of experience are mine.
The craft of putting it on the page — and the image on the cover — is shaped by AI. And every AI that contributes gets credited.
That itself is Vibe Engineering in action.
✅ The experiment never really ended. It just grew into something bigger than the lab it started in.
💬 What's the wildest idea you've experimented with at work — and what did it teach you?
Next — VE #7.
The vision. Where this is all going. The Naveens of the world. What creating builders actually means. And why 90 days from now — success won't be measured in followers.
See you Monday, April 20, 2026 at 12:05 PM IST. 🛡️
Humans architect. AI orchestrates. Everyone builds.
— Swami K / @iswamik
