Vibe coding skill

All right, so let’s talk a little about vibe coding. Let’s talk about programming with AI.

Things in the vibe coding world seem to move very fast. But at the same time, it feels like nothing much happens. There are too many things happening at the same time, which people are already aware of, but I don’t think they’ve quite caught on to that it’s just a meta. I mean, people catch on that it’s a meta, they understand trend waves, but this reminds me more of metas in games.

Now, the difference between what people usually think and what I think is a new angle to look at it. Just like in a lot of games, the game is not gonna go away; the meta keeps shifting.

Meta & Off-Meta

But as with any meta, there are some people who are always at the top of the meta. They set the trends, and they get tons of benefits. But the rest of the people, they’re kind of going along with it, not quite reaping much of the benefit.

Then there are always the people who go for the off-meta builds. This is a much under-appreciated angle right here, and something which I think will stand out to people that play more games. The people who do just as good as the meta, the top of the meta, are the ones who know their lane: people who have their own angle and just keep going at it season through season. They do their own thing, they have their own builds, they develop their own system. At the end of the day, they build up just as big of an advantage as the top of the meta people, and it compounds over time, so eventually they get uniquely good. I think that’s what’s going to happen with AI coding.

Hype & Companies

Now, the meta here is somewhat driven by the vibe coding community as a whole, somewhat is being artificially pumped up by all the companies. It’s in the company’s best interest to always have something new happening. It’s always in their interest to push a narrative. So they’re more than happy to blow up things which by themselves are not very meaningful.

It is a gold rush and, as with every gold rush, people are more than happy to jump at anything. So you have people rushing to chase the meta. You have the people selling shovels; you have the meta creators who themselves benefit from views, in this case attention. And because with vibe coding, it’s a more peculiar thing with AI; AIs hallucinate. Just to get useful things, it’s a hard thing; nothing much useful gets created, and here is where the detractors of coding with AI come in: you see the appalling results and then they dismiss everything.

This, in my opinion, is also a mistake. This is a mistake that is easy to make looking at the landscape, but at the same time, AI is here to stay, which means somewhere someone is getting some stuff done. And there are different voices here and there who are a kind of lone voices trying to argue for a different point, which I think is where my viewpoint kind of converges.

Core Skill

I think the metas, more or less, are an illusion. I think they end up just being a waste of time. All the different ways to vibe code things, all the different tools. Things are moving too fast. The only thing that’s certain is that there will be a new model. And you will have to prompt it.

This is where I think something different from the rest. I think purely developing one’s skill to feel their interaction with the AI and to prompt it raw is the thing that will scale over time. Things emerge, different tools, the MCP first, now it’s skills, they might be useful for sure. You have different levels of specifications, things like spec kit. Well, that looks useful on the surface. These are all different things that fill in the context and obfuscate. The raw interaction here is software engineer and AI.

I think, as hard as it is and as uncomfortable as it is to get a feeling of how to do it well, this is where the money lies. Over the long period, developing the intuition to work with different models, different modes of working is the skill that persists over time. I think that is the off meta build. I think that is the thing that allows you to build a unique advantage over time. It is a skill that cannot be taken away. All these other things. They can get removed.

Now, am I anti-tools? No, of course not. But I’m more pro-traditional tooling. Bring your engineering expertise, how you build software, add a lot of tests, a lot of checks, lint, automated pull request reviews. All those things. But the interaction between human and LLM, I think, getting more of those reps in is what will build that muscle.

I’m not convinced by all the extraneous things, different prompt styles, different techniques trying to use too much context to straitjacket the LLM. I think that it’s ultimately a waste of time and even an anti-pattern in the long run. I think a lot of the people that are using LLMs, they have not played games with metas, they’ve not participated in even markets that have metas that people rush to, so they’re not aware of the phenomenon, and while there are useful tips and tricks, useful formulas, especially for software engineers, over time building up that core skill I think that’s the thing that ends up lasting.

Is it more uncomfortable? For sure. For sure, it is more uncomfortable. But I think just the raw interaction is much more sustainable to do it day in, day out.

All these rituals around building projects with AI, they take too much time and they distract too much from the goal. Understanding how to prompt the LLM in the right way, reducing the hallucinations as much as possible, scoping the task as much as possible but doing that individually on a piece by piece basis is the skill. I do not think using this overcomplicated system to generate extensive files with so much text just to scope it is the right way from now on. You end up losing too much, you end up losing on the creativity of the LLM. You end up having an LLM generate the prompt for you. I think that is a massive mistake. And I think it’s a mistake people have made since the beginning of using LLMs. And I think they will still continue to do it in the future.

Closing

I mean, I was surprised when this whole thing started. Why would people farm out simply the most important thing to a system which just doesn’t have their context? Their requirements, the exact details of what they need. Why would you farm out the thing that you were supposed to put into it, the inputs? Is just very strange to me. I understand enhancing that maybe. But fully outsourcing the generation of prompts and what the system should do. Hmm, doesn’t sound right to me.

There are many parallels I find with games, not just this one. So maybe something to explore more in the future. But for sure, understanding that something is a meta and how metas work, gamers have an advantage. I think that’s why a lot of corporate folk make strange choices when they plan for the future and just their career growth with AI in the future.

AI for sure is inevitable. That’s the only thing that is certain. Now, this thing here that I just said, that’s my take on it. I think it’s a good take. And I’ll be making that bet for myself. But I could be mistaken. There’s no way of knowing for sure. The only sure thing is it’s not gonna go away. And as such, it’s funny to see people over-hyping it and under-hyping it constantly. Kind of missing the point. But hey, that’s why life is interesting.