Vibe coded plugins

[hey212]

I was testing Claude Code the other day and it failed on the very first try. I asked it for a slightly complex SQL query that should have contained a simple join, a compound left join, and a compound where clause. Its solution omitted part of the compound where clause and resulted in duplicate rows being returned. That's the kind of mistake that costs actual money if you deploy it.

Here's why Claude's mistake is interesting though - SQL is deterministic, like addition or subtraction. Given complete context, meaning a schema, a query, and a specific sequence of actions, there is one, and only one, possible outcome. If you were to type "2 + 2" into a calculator and get back "5", you'd throw the calculator away. You cannot understand SQL and make the mistake Claude made, unless maybe you were really tired, but Claude can't make that excuse.

LLMs are just insanely advanced pattern matchers that regurgitate things without understanding them. If you want to see just how far from possessing intelligence LLMs actually are, listen to the man in the funny shirt:

[Tiles]

Yeah, and that’s exactly why I say you still need real skills. A coder with a Dunning Kruger complex will most likely never finish the job. You need to understand what’s actually going on and recognize when the LLM starts heading in the wrong direction. And in best case fix problems that the LLM is not able to fix.

[legendCNCD]

But when the one whipping the LLM is not a coder, like not at all. Those plugins can stay the ... away from my installs

[Tiles]

And what if this plugin is better than anything else currently available?

For me, it comes down to experience, skill, and the final result, not the language or tool to get there. I know some people will probably want to stone me for saying that, but that is exactly what distinguishes a developer from a programmer. A developer will always look for the most efficient way to reach a working solution, and that does not necessarily involve writing code. Tools like Flowstone, HISE, or Reaktor demonstrate that quite clearly. That's quite a market still.

I develop a plugin with help of AI at the moment. Funnily enough, I actually have far fewer problems developing it with JUCE directly. My AI solutions completely failed at Reaktor and HISE.

In the end, getting to a working result at all already requires a fair amount of knowledge. As mentioned before, the idea that using AI equals clueless does not really hold up.

[stoopicus]

For someone who is already a software developer, AI tools offer ways to use your domain expertise in new and more effective ways. Most developers are figuring out to use it to automate the crap they hate doing themselves anyway.

As a developer, the amount of the job that is actually physically sitting there typing in code is small compared to coming to an understanding of systems, producing a design, and turning that design in to not only a working product but one that people would want to use.

AI tools can accelerate all these things, but still require the domain expertise. It works much better to work with the agents to approach problems iteratively and factor them in to appropriately sized chunks and produce ways to test those bits well in isolation and focus on reuse and... well, all these things things we already do. Things that were good practices before are still so now. It's just shifting the focus more towards architecting and orchestrating and away from the more entry level skills like simply coding.

AI tools are also still spectacularly bad at one-shot tasks where you leave all the domain expertise up to them. So sure, there will be lots of wankery with people thinking "I'm just going to tell the AI to make me the Next Big Plugin!" and then producing laughable crap. Who cares?

[ghettosynth]

For someone who is already a software developer, AI tools offer ways to use your domain expertise in new and more effective ways. Most developers are figuring out to use it to automate the crap they hate doing themselves anyway.

As a developer, the amount of the job that is actually physically sitting there typing in code is small compared to coming to an understanding of systems, producing a design, and turning that design in to not only a working product but one that people would want to use.

AI tools can accelerate all these things, but still require the domain expertise. It works much better to work with the agents to approach problems iteratively and factor them in to appropriately sized chunks and produce ways to test those bits well in isolation and focus on reuse and... well, all these things things we already do. Things that were good practices before are still so now. It's just shifting the focus more towards architecting and orchestrating and away from the more entry level skills like simply coding.

AI tools are also still spectacularly bad at one-shot tasks where you leave all the domain expertise up to them. So sure, there will be lots of wankery with people thinking "I'm just going to tell the AI to make me the Next Big Plugin!" and then producing laughable crap. Who cares?

100%.