The Famous Garfish's Tank

Vibe coding

Vibe coding has become the expression du jour for letting large language models (LLMs) write code for you. It lives on a spectrum between the agentic, fully integrated IDE that runs commands, reads files and writes new ones at the merest suggestion you want it to do something and glorified autocomplete.

Please, I had a long day, stop calling it that

I know, as a self-identified grump and hater, the term ‘vibe coding’ makes me die a little inside every time I read it or hear it used unironically. But saying it is fun. It’s one of those terms that the youths, the rappers (same thing) or the nerds would allegedly use as slang, but really, it’s mostly your streamer’s favourite streamer and other millennials using it compulsively (‘ironically’). Like the ancient ‘YOLO’, ‘pwned’, ‘gank’, ‘1337’ and ‘swag’, or the more recent ‘rizz’, ‘sigma’ and ‘sus’. Of course ‘vibe coding’ is in the center of the Youtube, business and gen Z slang bubbles, which makes it a perfectly disgusting thing to say while eagerly checking the people around you for nervous eye twitching.

To avoid it, let’s call the practice of letting an LLM agent program for you, preferably in a language you don’t know anything about, Claude sitting, or more generically bot sitting.

How I learned to stop worrying and love the bot

As I wrote previously, I don’t think letting LLMs write code for you is as problematic, ethically speaking, as other forms of generative AI. Is it actually good, though? I have used various models in various forms and various languages for the last years and I have changed my mind about it a bit. In the beginning, I was very negative, whenever someone tried to convince me that absolute novices could produce useful code using chatGPT. I work in an academic setting and part of my job is to program in R at an intermediate level. As Tim Scott puts it in aRrgh: a newcomer’s (angry) guide to R:

R is a shockingly dreadful language for an exceptionally useful data analysis environment. The more you learn about the R language, the worse it will feel. The development environment suffers from literally decades of accretion of stupid hacks from a community containing, to a first-order approximation, zero software engineers.

Tim has since repented and added a disclaimer to his guide, but he is not too wrong about R and its developers. So I thought my early bad experiences with letting chatGPT cook up R code was a combination of those models being fundamentally unable to reliably code and R being too obscure and the data the models were trained on being full of bad code. For me, this changed around the time OpenAI released chatGPT 4. Not because it produced better R code, but because it became bad at doing the one thing I had used 3.5 for routinely: writing function documentation. So I changed to Anthropic’s Claude and was blown away. Claude just wrote documentation without being as chatty or wrong as chatGPT was. Eventually, I started trusting Claude enough to ask him for actual code, first in R, then Python and eventually everything else. So Claude is just better? No, he was just my first love.

If Copilot is moving fish, who’s flying the plane?

With this website project I have fully become a bot sitter. A relative novice at HTML and CSS, just prompting away and watching the code write itself. At the best times it’s like magic. From a simple instruction with typos and non-technical language, the agent looks around himself and spins up a whole host of pages that all look like real websites link to each other and he turns to you and smugly asks “anything else, boss?”

Most of the time, it is not like that. Most of the time, something just doesn’t work and you keep prompting variations of the same sentence in increasing desperation, expecting a different result this time. Every time the bot finds a new possible reason for what could be wrong and, getting supposedly agitated, it adds more and more code that looks like it should not be there. Even worse, sometimes it suddenly works and since you don’t have a clue what any of it means, you are afraid to touch any of the mess that the two of you created in a teenage-house-party binge of prompts and hundreds of lines of unread answers and sweat and tears and server usage. Your parents never come home to look at you full of disappointment, but there’s an uneasy feeling that you shouldn’t have done that.

Unlike the name of a popular AI autocomplete implementation suggests, the bot you’re bot sitting is not a copilot. They’re the pilot and will fly the plane with unwaning enthusiasm. You are not expected to know how to fly, you’re expected to babysit the pilot, because they’re not actually that good at it either. If you want to not crash, you either have to stay in the air indefinitely, get lucky, or learn enough about aviation from the manual to guide them through the steps and, maybe, eventually, you will have accidentally learned a thing or two about planes.

Here’s an example where I asked a Gemini-driven agent in cursor to move a fish on the landing page microblog feed a bit further below the box, this ‘reasoning’ process took 56 seconds, feel free to not read it, I certainly didn’t:

Anyway, so I moved the fish myself

So bot sitting is not a magic bullet to let novices code, I think. Fundamentally, it’s a great way to get started quickly, there’s no hello world anymore, there’s a whole week of struggling worth of code generated in seconds. The start will most often look and work very well for no effort on your side. But slowly, you will have to understand what is actually happening and how it happens in order to fix the small and large issues the bot created in its eternal effort to make sense of the next word given its limited context window. You will have to grow up fast, the bot won’t. You will come home from school, looking at the mess you both made, you will be disappointed in the both of you and do better next time. Maybe.

Had I written the CSS that positions the fish myself, I would have immediately been able to move the fish without any reasoning. This way, I had to scroll through the whole jungle of style Claude had made for my page, find the fish and try three different settings to get it below the box. Gemini had moved the fish into the box after finishing its deep thought. But I learned something about organising the CSS file, relative and absolute positioning. For me, that was a more fun way to learn than to do yet another “hello world” and lose interest after that. I was vibing.