One of the things about software development is that it is still inherently expensive to do; well at least, to do it well. There have been countless attempts to reduce the cost of software development leading to innovation in terms of both tooling and processes. In this post I’m going to discuss the more recent trend around Vibe Coding and specifically how it’s being applied to some of the AI-first tooling that’s available. The aim is to answer the question as to whether Vibe Coding will enable us to build better, or cheaper, software.
I want to start by saying that, as with most tools, technologies, processes etc, there’s going to be times when Vibe Coding is the right tool for the job (I use the term “tool” here loosely here as it may not be a specific tool, perhaps instead a process, or lack thereof). Furthermore, you shouldn’t read into this post any judgement on whether Vibe Coding is a good or bad practice. At the end of the day, whether Vibe Coding is right or wrong, should be determined based on the achieved outcome in terms of time, cost and quality of the output.
Let’s back up a little and look at what we mean by Vibe Coding. For a quick summary we can track Vibe Coding from being an Internet meme through to being a more widely recognized practice in software development:
🌐 Internet meme → 💻 Indie dev/streamer slang → 🤖 AI-assisted era → 🎤 Recognized as a legit creative practice.
| Phase | Description | Reference |
|---|---|---|
| 1. Coined by Karpathy | February 2, 2025 post on X describing the concept | (Know Your Meme, The Daily Dot) |
| 2. Meme Spread & Community Jokes | Rapid memetic adoption across X, Reddit, TikTok, Hacker News | (smartlab.at, blog.vibecoding-hub.com, Reddit) |
| 3. AI-Assisted Coding Practice | Shift to prompt-driven workflows, minimal code scrutiny | (Wikipedia, Pluralsight, GitHub) |
| 4. Cultural Legitimization | Acknowledged in tech media, investment circles; recognized as creative method | (Wikipedia, Forbes, Business Insider) |
I have to say that I truly appreciate the honesty of the Merriam-Webster definition of vibe coding:
Writing computer code in a somewhat careless fashion, with AI assistance
Merriam-Webster
In Practice
One of the hip trends online is the slew of low or no-code online app builders such as Lovable, TemoLabs and co.dev. Most of these start with a singular prompt, encouraging you to give a one or two sentence description of the app that you want to build. The builder then goes off and builds a very rudimentary application that looks very run-of-the-mill.
The experience of then interacting with the builder site in order to coerce it to give you the application you want can be both rewarding and somewhat infuriating. The ability to create a fully-fledged application from a mere sentence or two is simply astounding, if you compare to where we were even 2 years ago in the world or low-code/no-code solutions. It can be really rewarding to have an idea, generate a simple, yet working, prototype and be able to validate the idea in minutes, rather than days, weeks or months.
On the flipside, sometimes the model can really struggle with what might seem like almost basic issues. This could be a simple compile error that for whatever reason the model can’t rectify without some intervention. As the models and site integration of the models improves this will get better but inevitably there will be compile errors, logic errors and complexity that is beyond the scope that can be resolve via a simple prompt-based interaction.
There’s also the cost aspect of these builder sites – they significantly reduce the cost of creating applications but at the expense of building something that has a strong architecture and is able to mature and grow with the requirements of the application. Furthermore, once you get beyond the included tokens, you need to start paying for tokens in order to keep feeding the AI algorithm. The more complexity in the app, the more likely the model will yield build errors, which in turn mean more tokens consumed to fix the errors the model created (yeh, there’s no “you broke it, so fix it for free” when it comes to AI).
If we consider the app builder sites as the completely hands-off approach, where you’re simply interacting via a prompt with minimal interaction with code, the other trend is to interact with AI models using IDE or console tools, like Cursor, Windsurf, GitHub Copilot (VS Code and/or Visual Studio), Codex, Warp etc. Most of these tools allow you to control which AI model you want to use and the type of interaction you want to have with the model. For example, this illustration is from the Chat pane inside VS Code, allowing you to configure the mode (Agent) and model (Claude Sonnet 4).
What you select for the mode really determines the level of actual coding you want to be doing. If you’re deep in the trenches writing a lot of code, the Ask mode is great for suggesting ways to code a method, analysing existing code, providing hints and information on how to use an API etc. Of course, if you want the model to be more involved, you can switch to Edit mode, in which case the model will start to apply changes, and potentially changes across a number of files. Lastly, if you want to be really hands-off you can set the mode to Agent. Agent mode is great for where you want to hand off a coding task and simply be involved in the reviewing phase. In the recent release of Codex, this has gone one step further, making it easy to dispatch whole coding tasks into the cloud to be done in an asynchronous manner, freeing you up to continue with other tasks.
Cost Cutting
The opening premise of this post was to consider vibe coding in the context of building better, and cheaper, software. In the preceding section we looked at two of the approaches to vibe coding, app builders and AI enabled tooling, and I would argue that there’s a mix of pros and cons with both approaches.
If we look at the app builders there is definitely a cost saving when it comes to the ability to generate quick, working, prototypes in order to validate ideas. It’s questionable as to whether these can easily be used as the basis for quality production-ready application without substantial investment in developers to review, refactor and stabilize the codebase.
On the flipside, AI enabled tooling can offer a mixed bag of success. This is a much larger topic that’s really still in its infancy regarding the optimum way to interact with AI models but enough to say that even if you were to occasionally use the AI models in Ask mode, you’d still derive some benefit from the speed and knowledge that the AI model brings when it offers up solutions.
In Summary
I think it’s too early for companies to be making significant reductions in their development teams. Rather they should be investing in internal knowledge/research into the ways that AI can assist the existing development team to be more productive. Teams that can adopt AI tools and leverage agent mode will surely be able to deliver more features, quicker, and with more stability.