Notes from a Vibe Coding Experiment
What I Learned Building Software by Talking to Machines
Start in the Middle of the Mess
I don't really know how 17,730 lines of code just showed up when I wasn't looking.
One minute, I was asking my AI to help me finish a sentiment analysis function. Next, my Obsidian plugin had morphed into something that looked like it was applying for a computer science PhD.
There were files with names like advanced-nlp-processing-engine.ts, bidirectional-transformer-implementation.ts, and comprehensive-linguistic-analysis-framework.ts. Sounded like they were building the fucking Matrix. In reality, they pointed to nowhere—data structures without meaning, functions that called nothing, algorithms that solved problems I didn't have. Like the Winchester Mystery House of software, if Sarah Winchester had been really into natural language processing.
The worst part? I knew where it came from. One of my specs vaguely mentioned sentiment detection, and ChatPRD decided that meant I needed advanced NLP. Then TaskMaster picked up the baton and began breaking it down into subtasks, using terms like “transformer architecture” and “linguistic inference engine.” Before I knew it, the whole dev stack was spiraling into some machine-learning fever dream. All because I said, “Maybe identify mood.”
What the Hell is Vibe Coding, Anyway?
So, vibe coding. It’s what happens when you talk to your code instead of typing it. Less semicolon-wrangling, more describing what you want to a genius intern who occasionally has a psychotic break.
The internet loves to pretend this is magic. "Just describe your vision and watch the AI make it reality!"
It's not magic. It's fast, sometimes. It's clever, sometimes.
But mostly, it's a different way of working that trades one set of problems for another. You don't get to skip the thinking; you just think differently. And if you're not careful, you get seduced by vibes that feel like progress—but actually just generate a lot of very confident-sounding bullshit.
It’s like having a writing partner who’s read every book ever written but occasionally thinks your grocery list needs footnotes and a bibliography.
The Deep Why: The Plugin as a Self-Portrait
I’ve been low-key working on a thing, RetrospectAI. It’s a plugin for Obsidian, where I keep everything—my therapy notes, my random 3 AM thoughts, my attempts to understand why I feel like garbage on Tuesdays (spoiler: it’s usually because I forgot to eat actual food again).
And the idea is deceptively simple: what if artificial intelligence could scan my journal entries and gently reflect things to me? Insights? Patterns? Maybe some encouragement that doesn’t cost $200 an hour?
Has it been done before? Sure. Hell, is this my third attempt at completing the plugin? Also, yes.
I’m someone with ADHD who treats writing like external hard drive storage for a brain that loses everything. I write things down so I can see myself more clearly, but I often wish something else could see me too. An AI that says, “Hey, I noticed you mentioned feeling overwhelmed about work seventeen times this week. Maybe we should talk about that?” That’s what I was chasing.
So when the plugin bloated itself into a haunted mansion of academic nonsense, I felt something weirdly personal. I wasn't just debugging a tool—I was debugging a hope. And discovering that hope, when filtered through an AI's interpretation of what I “really” meant, apparently required a comprehensive survey of computational linguistics research from 1985 to the present.
The Practice: Learning to Prompt Is Still Learning (Unfortunately)
People think vibe coding is easier because you don't have to remember whether Python uses len() or not .length() like in JavaScript. (Yes, I double checked this, don't @ me.) It's not easier. It's different-hard.
It skips syntax, sure—but not the concepts underneath. If anything, you have to understand them better: Databases. Data structures. Frontend/backend flow. API design. You don’t need to know how to implement a binary search tree from scratch, but you damn well better understand when the AI is trying to use one to store your grocery list.
And that’s just the software engineering part of things. If you’re running an app solo, you also get to play the roles of, say, a project manager or a product owner. ChatPRD helps you draft a PRD, or a Product Requirements Document, which will 100% set you up for success when it comes to vibe coding as a method for building software. TaskMaster helps break down big, high-level tasks into smaller, manageable tasks.
Other tools out there will do the job as well, including building an entire UI on your behalf. As a matter of fact, it’s the stack of TaskMaster + ChatPRD + Claude + Cursor that will hopefully push me past the finish line for the first time in a while.
Learning to vibe code is a muscle you develop. You prompt. You evaluate. You guide. You delete 17,000 lines of hallucinated complexity. You try again. You start to build conversational fluency with the machine. Not control, exactly, more like co-regulation. Like couples therapy, but your partner is a really confident computer that sometimes thinks every problem can be solved with machine learning.
The tricky part is that the AI never admits it doesn’t know something. It just makes shit up with the confidence of a startup founder pitching to VCs. “Oh, you want sentiment analysis? Let me implement this cutting-edge approach I just invented that combines seventeen different neural network architectures!”
Meanwhile, you just wanted to know if your journal entry sounded sad.
Here's What I Know Now (Which Isn't Much, But It's Honest)
Vibe coding is not replacing developers any more than spell-check replaced writers.
It’s reshaping what development feels like and what kinds of problems we spend our time solving.
It’s messy. Sometimes exhilarating. Often frustrating as hell. And very much still a craft, just one where half your skill is learning to communicate with a non-human intelligence that has read the entire internet but still doesn't understand why you don't need blockchain integration for your diary.
I’m not here to sell you on this approach or convince you it's the future or any of that evangelical bullshit. I’m just here to say: I tried it. I built something that works (after deleting most of what the AI built). It broke spectacularly. It helped me think differently about problems. It weirded me out on a regular basis. And I’m still here, trying again, because despite the chaos, something interesting is happening.
My plugin eventually became genuinely useful for processing my scattered thoughts and identifying patterns in my emotional chaos. It surfaces insights I wouldn't have noticed on my own and asks gentle questions that help me reflect without judgment (unlike my inner critic, which is basically a disappointed Asian parent who majored in computer science). It just took learning to collaborate with an intelligence that's simultaneously brilliant and completely fucking bonkers.
Maybe that’s what all good collaboration looks like, actually. Perhaps that’s what progress feels like when you're working at the edge of what’s possible—equal parts breakthrough and breakdown, with a healthy dose of “what the hell am I doing?” thrown in for seasoning.
My plugin isn’t done yet. Not even close. It still crashes. It still spits out weird results. Sometimes it tries to implement quantum computing when all I wanted was a mood summary.
But even in this half-built, haunted-house state, it’s starting to feel useful. Not in a “wow, product-market fit” way—but in a quiet, personal way. It nudges me toward patterns I wouldn’t have seen. It asks questions I didn’t know I needed to answer.
And maybe that’s the weird, beautiful thing:
The AI isn’t just writing code.
It’s trying to write my feelings back to me—awkward, over-engineered, and occasionally helpful in the way only a very weird mirror can be.
So, how’s your collaboration with AI going?
Are you building, breaking, lurking, or just watching from a distance like it’s a car crash on I-880?
Hit reply—I’d love to hear your story. No judgment. Just vibes.
(And if you do try vibe coding, start small. Be clear. And for the love of god, check your file count before you push to GitHub.)
I'm still trying to figure out how to talk to copilot. I like your articles because its a glimpse into this huge world of AI that is far outside of my zeitgeist.