How do we kill vibe coding and AI slop?
Rob Patro October 23, 2026
Has Rob reversed his thoughts on the capabilities of AI tools completely?
No, I have not. Rather, I’m using the title here as a rhetorical hook. Specifically, I’d like to “kill” vibe coding and AI slop in two senses: stop producing the careless, low-quality content these terms describe, and stop applying the same labels to careful, expert-driven work solely because AI tools were used in its creation. I’d like to address these two issues separately.
It’s time to stop vibe coding and producing AI slop
The first sense in which these terms are used is the “historical” one. Vibe coding is a term used to refer to AI-driven development of software in which one “describes a project or task in natural languages to a large language model (LLM), which generates source code automatically.” Likewise, according to Wikipedia, “Vibe coding may involve accepting AI-generated code without thorough review of the output, instead relying on results and follow-up prompts to guide changes.” It has been argued that this process tends to produce software that can be categorized as AI slop, which is described as “digital content made with generative artificial intelligence that is perceived as lacking in effort, quality, or meaning, and usually produced in high volume to gain advantage, to earn money, or to deceive people.” (initially used with respect to AI images or videos, but later expanded to refer to all manner of AI-generated content).
Now that we’ve had some time to see the quantity and quality of content that generative AI models are capable of producing with minimal effort or oversight, it seems quite clear that we would almost always be better off without such content than with it. Purely vibe-coded software often presents as poorly designed, with frustrating bugs and quirks. Moreover, given that this software often takes very little effort to create, the creators often have very little interest in maintaining it, fixing these bugs, responding to user feedback, or improving the quality of this software over time.
Of course, these applications can sometimes be useful — when one needs to create a simple one-off application to perform a task or organize some data, for example. Likewise, for the simplest and most common use cases (where many established alternatives likely exist, and have been observed by the model during training), even the simplest prompts can produce satisfactory results. However, for even moderately complex applications or software tools, weak and low-effort prompts tend to produce software that can be meaningfully defined as “vibe-coded slop”. However, to claim or believe that the quality of such output is an inherent characteristic or limitation of the models is to misunderstand how they may be used as tools.
I’ll also point out a corollary here, which is related to but distinct from the issue of low-effort prompting. Low-skill prompting — where the prompter does not well understand what they are looking to create, how it should work, or how it might be built — is also likely to result in poor results. It has been demonstrated, of course, that intelligent users putting in substantial effort can obtain good results, even if their expertise is not in software development, but I think we have reason to view this more as the exception than the rule, at least for now. The gap in the quality of software created by people who have “traditional” software engineering and algorithm design experience and those who don’t seems substantial. Given how efficient and powerful these models can be (see below), a main concern I have is actually that, rather than enable novice engineers to operate at the level of senior engineers, AI will instead create an even wider gap between the capabilities and productivity of novices and those with substantial experience (across many disciplines and not just software engineering), making it increasingly difficult for the novices to catch up.
It’s time to stop calling agentic software development vibe coding, and AI-assisted software “slop”
First, I want to make this claim and open this discussion from a technical perspective. I have heard and read many coherent and substantive ethical critiques of generative AI models, how they are trained, and how they are used. That is a critically important conversation to have, and I have engaged in it to some degree and would like to do so more. However, here, I’d like to focus on the technical merits of high-effort, expert-driven, AI-assisted software and tool development.
After engaging with several people who have made such claims, and seeing how the term is often used, I can only conclude that the people who persist in calling this software “slop” solely because AI tools were used in its creation are either (a) disingenuous (attempting to apply this label while knowing better because they take issue with the technology itself); (b) intellectually lazy (unwilling to engage honestly with the technology and repeating what they’ve heard from others or what they recall from trying predecessor technology 1-2 years ago or more); or (c) unqualified to judge the output they are critiquing. It seems “inconceivable” that someone with adequate skill and experience to make such judgements might look at the result of an expert using these tools, expending effort to focus on the high-level design and ensure the right low-level details, and proceeding via a careful process, and deem the result “slop” solely because AI tools were used in its creation. I am certain that we will see numerous studies in the future that quantify and evaluate how such software operates, endures, and performs over the long haul. However, from the perspective of having built and used such software, the anecdotal evidence is overwhelming that the quality and robustness can be extremely high, just as high as if the entire project was made from artisanal, hand-crafted code, sprinkled with carefully considered SIMD kernels and just the right amount of inline assembly to eke out every last drop of performance.
Moreover, these tools provide an ability to plan, design, and iterate on solutions faster and in more detail than would otherwise be possible, which is critical in iterative improvement. Likewise, their ability to optimize software, from high-level details like data structure mismatch to lower-level details like allocation patterns, to important considerations like cache access patterns and memory packing, is excellent. Perhaps unsurprisingly, the workflow for doing this looks much like what a software engineer might do. Compile your program with debug symbols, record a run under different workloads with perf, and analyze the resulting breakdown of where time is spent. Track the cause of unnecessary allocations, unexpectedly high instruction counts, and cache misses. Even aspects of software development like maintaining code hygiene can be made easier, more uniform, and more universal with agentic development.
Again, the claim here is not that software being written by or with AI is automatically as good as or better than software written by a human developer. Rather, the claim is that when agentic tools are used by an experienced developer, and that developer commits substantial effort to the project, the resulting quality can be very high (of the same order as it would be had it been developed without agentic tools). To call such output “slop”, solely because AI was involved, therefore, is both to expose one’s own ignorance and to devalue the term to the point of meaninglessness. Likewise, the process used to obtain such output is not reasonably described as “vibe coding”, and there’s no reason to refer to it as such.
It is time…
It is time for us to move on from the novelty of cookie clicker AI capabilities. The internet is already filled enough with vibe-coded software and AI slop. That content almost always detracts rather than adds value; it is worse than making nothing at all. I sincerely hope that we can move on from that flood of low-human-effort content as quickly as possible.
Also, it is time to stop carelessly or dishonestly calling a development process “vibe coding” and its output “slop” solely because agents were involved. If one has reasons to oppose the use of AI in a specific circumstance, or more generally in a certain field or type of creative endeavor, then just say this. However, cynically and disparagingly labeling as “slop” content that was created with great human effort and expert understanding, simply because you disapprove of the process or the tools by which it came to exist, is at best intellectually lazy, and at worst intellectually dishonest.
In both meanings of the terms, it is time to kill vibe coding and AI slop.
Comments