Three Months with Claude: Vibe Coding or Vibe Creating?

Author: Hoc Tro and Claude Code (Anthropic AI)

Division of labor:

• Hoc Tro: Originating ideas, directing content, cultural knowledge, contributing to drafts, fact-checking, final review
• Claude (Anthropic AI): Analyzing source material, drafting text, maintaining voice and style

Editor-in-chief: Hoc Tro

---

I. "Vibe Coding" or "Vibe Wandering"?

Not long ago I read an English-language article with a fairly sensational headline: "These coders want AI to take their jobs." The author, Ariana Aspuru, describes a phenomenon spreading through the software engineering world in America: people are using AI tools like Claude Code, Codex, or Gemini to write code, run tests, and push straight to production — without ever typing a single line themselves. Andrej Karpathy, co-founder of OpenAI, coined the term "vibe coding" for this phenomenon — meaning you surrender completely to the flow, trust the machine, and forget that code even exists. The article tells stories of small startups working twenty times faster, of Amazon and Google moving more cautiously but still seeing productivity rise by ten percent, of developers who feel both excited and uneasy at the prospect of being replaced by the very tools they use.

I finished reading and sat there laughing to myself. Not because the story was funny, but because I realized I had been doing exactly the same thing — with one difference: over these past three months, I haven't been using Claude to write code. I've been using it to write essays, analyze music, and research song lyrics. Not because I don't know how to code — I'm a programmer, I write code every day. But when I saw how sharply Claude could analyze things, I decided to let it do first what it does best. As for coding — I'll bring Claude into that work too, no question about it. Over these past three months, I've used Claude to do things I never would have attempted on my own: writing dozens of long musical research pieces running into the thousands of words, analyzing Italian and Japanese lyrics, planning the migration of legacy engineering software to modern systems, drafting hundreds of AI application scenarios for the oil and gas engineering industry. If Karpathy calls his approach "vibe coding," then I'll call mine "vibe creating" — surrendering to the creative flow, trusting Claude as a genuine collaborator, and letting ideas pour out without stopping.

What strikes me is that the article worries about the future of programmers. And here I am — a programmer and engineer who also happens to research music — sitting down to record my thoughts after three months of collaboration with Claude, and I feel something completely different: not anxious, not afraid of being replaced, but simply astonished at what I've managed to do.

---

II. The Biggest Project — Researching the Music of Phạm Duy

If I had to name the single project that consumed the most time and care over these past three months, it would be the Phạm Duy music research project. I've been tied to the website phamduy.com for many decades now — back when it was still kycon.com in the days when Vietnamese on the internet had no Unicode, then phamduy2000.com, then phamduy2010.com, and now the current phamduy.com in its second iteration, managed by NH. All those years, I had always wanted to write serious research pieces about his music — not the kind of breezy introductions you find everywhere, but real analysis that goes deep into each song, each creative period, each distinctive musical technique he employed. But working alone, with limited resources and not enough hours in the day, there was only so much I could do.

Claude changed that equation entirely. Over three months, Claude and I completed twenty-eight research pieces, ranging from major works like Tình Ca, Con Đường Cái Quan, Mẹ Việt Nam, Đạo Ca, and Rong Ca, to more specialized comparative analyses like Comparing the Language of Phạm Duy and Trịnh Công Sơn and Phạm Duy and the French Chanson. Our working process gradually found its rhythm: I would set the direction and provide source material — memoir passages and text corpora — Claude would draft the piece section by section, and then I would read it back, revise, add and subtract according to my own sense of what was true. Each piece was saved as a separate file, then assembled into the final article using a Python script — never letting Claude re-read everything and rewrite it from scratch, because experience taught me the machine would quietly "summarize" without knowing it was doing so, and all the precious details would be lost.

But not everything went smoothly from the start. There is one painful lesson I want to share directly here: Claude can fabricate folk poetry. During one round of reviewing the complete set of 132 pieces in the 100 Tình Khúc collection, we discovered that several folk verses had been cited with great precision — they sounded absolutely authentic, perfectly in keeping with the oral tradition — but had no verifiable source whatsoever. We called it "hallucinated folklore": the machine inventing proverbs and folk verses that looked real in order to fill gaps in the argument. From that point on, I adopted an iron rule: never quote folk poetry or proverbs without a clearly printed source. Better to say less and be right than to say more and be false. This is a lesson not just for me, but for anyone experimenting with human-machine collaboration in academic work: rigor is something you cannot delegate to the machine.

All the pieces in this project carry two names: Hoc Tro and Claude Code (Anthropic AI). I have no hesitation listing Claude's name, because I believe it is the honest thing to do. Claude didn't merely "type for me" — Claude genuinely participated in the analytical process, drawing connections I might not have thought of, suggesting perspectives from Western musicology that I might not have had time to explore on my own. But the one who made the final calls, the one who knew where the real story lay, the one who understood Vietnamese culture and the Vietnamese people — that was still me.

---

III. From Naples to Tokyo — Analyzing Foreign Music

Alongside the Phạm Duy project, I also used Claude to explore international music in a way I never could have managed alone. I'll say it plainly: I've loved Italian and Japanese music for a long time, but my language skills only go so far. With Claude, I was able to go deep into lyric analysis, understand the circumstances in which each song was written, and produce serious research pieces — not Wikipedia summaries, but the kind of analysis that actually means something.

The story I love most from this section is about Peppino di Capri and his song Un Grande Amore e Niente Più. He was born in 1939 on the island of Capri and is eighty-six years old as of this writing, and the song is attached to a beautiful piece of legend: Franco Califano — one of the great figures of Italian pop — stayed up five consecutive nights writing the lyrics, because he was absolutely determined to find a closing line worthy of Peppino's melody. That closing line, when he finally found it, was so simple it was almost impossible to believe — but it was the kind of simplicity that only a genius can write. When Claude helped me analyze the harmonic shift from E-flat to E major at the end of the song, I sat and listened to it again and felt I understood it as I never had before.

Then there was Annalisa Minetti with Senza Te O Con Te from Sanremo 1998 — a blind woman who had been Miss Italy in 1997 and went on to become a Paralympic athlete, writing a song with the paradoxical title "Without You or With You" that was really the voice of her own heart about the life she lived without sight. Riccardo Fogli with Storie di Tutti i Giorni — the story of Mario and Maria, two ordinary people, and the way the composer Guido Morra wove it into a song about the quiet dreams of ordinary lives. Each song is a world unto itself, and Claude helped me open doors I had only ever stood outside and looked through.

With Japanese music, I entered the world of Mika Nakashima — the actress born in 1983, best known for playing NANA in the film adaptation, with a voice thin as silk that somehow strikes straight at the heart. We analyzed thirteen songs, from Yuki no Hana to Hatsukoi, and I came to understand that Japanese music has its own very particular way of approaching loneliness — not mournful the way Vietnamese music can be, not dramatic the way Italian music tends to be, but something like a very gentle, very accepting silence.

---

IV. The AutoCAD Menu and the Engineering Work

Now I'll turn to a completely different area — one that, given the title, might make you wonder: what does this have to do with Phạm Duy's music? The answer is: nothing at all. But that is exactly the point I want to make about Claude — that this tool is flexible enough to follow me from the world of art into the world of engineering without missing a step.

I'm an engineer. In my day-to-day work, I use AutoCAD and various other engineering applications. For a long time I'd had a collection of old tools written in SmartSketch — a technical drawing application by Intergraph that most people today no longer remember. Those tools were extremely useful, but the original software had lost support long ago. My idea was to migrate them to AutoCAD, rewriting them in VB.NET to run in a modern environment. On my own, I never had the time to sit down and read through the API documentation, compare equivalent functions across the two systems, or plan out each phase of the migration. Claude helped me do all of that — not by writing the code for me and pushing it straight to a server in the "vibe coding" style the article described, but by sitting down with me to plan the work, drafting documentation, explaining the differences between the two API systems, and writing guides for me to refer back to later.

There was also a larger project that I'm very proud of: a body of documentation called AI in the Engineering Industry, comprising 383 specific AI application scenarios for EPCM companies — that is, engineering, procurement, construction, and management consultancies operating in oil and gas, pharmaceuticals, and urban infrastructure. Each scenario was written to a fixed structure: what is the problem, how did people handle it before AI, how does AI help now, what are the steps to implement it, and how does AI verify its own results. All 383 scenarios, spanning civil engineering, structural, process, cybersecurity for OT systems, drone inspection, and digital twins. This is the kind of work that used to require hiring a team of consultants — and even then, you might not get the depth and breadth of coverage that we produced.

---

V. Learning About Claude Through Claude

There is something I haven't mentioned yet, and it is interesting at a rather meta level: I used Claude to learn about Claude itself. In my Converting Transcripts folder I have dozens of interviews and presentations from people at Anthropic — Boris Cherny, Dario Amodei, and many others — talking about how Claude Code works, about the design philosophy behind it, about features most users never discover. I asked Claude to transform those long, sometimes messy transcripts into clean, readable pieces organized by chapter with clear headings, correcting the recognition errors that turned "Claude" into "Cloud" and "Anthropic" into "Enthropic," and formatting everything as HTML with color-coded speakers.

This turned out to be a satisfying loop that I hadn't anticipated when I started: learning about Claude by using Claude to process articles that talk about Claude. And in doing so, I came to understand that the team at Anthropic built this tool with a very clear philosophy — not to replace people, but to amplify what people can do. Dario Amodei calls it "human-centered AI." Boris Cherny has talked about how Claude Code was built to serve people doing real work, not to put on a performance.

I believe in that philosophy, because I have lived it over the past three months.

---

VI. Closing — Small Victories Arriving Like a Waterfall

The article "These coders want AI to take their jobs" ends with an interesting observation: in the old days, a programmer's joy came from the small victories — when a piece of code ran correctly, when a bug was finally squashed. Now, because AI handles all of that, those small victories are gone — but in their place, "the big wins come in like a waterfall and it's intoxicating." I read that line and felt it described my own experience exactly.

Over these past three months, I've had many moments like that. The moment of looking at a six-thousand-word research piece on the screen and knowing it was right — not just factually right, but right in its voice, right in its feeling, right in the way it told the story of Phạm Duy. The moment of listening again to Un Grande Amore e Niente Più after finishing the analysis and finding I was hearing the song in an entirely different way. The moment of looking at a table of 383 AI application scenarios and thinking: this used to take a whole team several months.

But I have to be honest: not everything has been a victory. There are times Claude is overconfident, inventing details that sound entirely convincing but turn out to be wrong. There are times the writing drifts toward a Northern Vietnamese register, or the tone marks disappear entirely when the context window is running low. There are times I've had to repeat the same instruction over and over, thinking I'd already made it perfectly clear. This is the real side of human-machine collaboration that few people talk about — it isn't always smooth, it doesn't always "vibe" in the true sense of the word.

The most important thing I've learned is this: AI cannot replace genuine knowledge. Claude does not know Phạm Duy the way I do. Claude does not know Peppino di Capri the way someone who has been listening to his music since the 1960s does. Claude does not know what it feels like to sit at a drafting table and understand from the inside why an old tool was designed the way it was. Those things I have to bring with me. Claude only helps me turn them into words faster, more completely, and with greater coherence.

The article asks: what will programming look like in five years? I don't know. But I know that over these past three months, what changed was not my work itself — it was the scale of what I could accomplish within the same stretch of time. Before, I might write one music research piece every few weeks. Now I can write twenty-eight pieces in three months — and each one still carries my mark, still speaks in my voice, still stands as my work. I just no longer have to do it alone.

Perhaps that is the real answer to the question the article was asking. It isn't that AI takes people's work away from them — it's that AI helps people do the work they never dared to dream of doing on their own.

---

Completed March 2026