The Software Engineers Paid to Fix Vibe Coded Messes

Typical coding LLM issues:

Hallucinations

Context limits

Lack of test coverage and testing-based workflow

Lack of actual docs

Lack of a spec

Great README; cool emoji

    Lack of actual docs
    Lack of a spec

>If the resulting software is so poor you need to hire a human specialist software engineer to come in and rewrite the vibe coded software, it defeats the entire purpose.

I don't think this is entirely true. In a lot of cases vibe coding something can be a good way to prototype something and see how users respond. Obviously don't do it for something where security is a concern, but that vibe-coded skin cancer recognition quiz that was on the front page the other day is a good example.

Like Red Teams for InfoSec, reliability teams meet developers. Not new, but keep pumping Gig Culture/the fad, I guess.

Well, I'm sure we've all seen code produced by human developers that is 10x worse than what my Claude Code produces (certainly I have), so let's be real. And it's improving scary fast.

I get two types of merge requests nowadays. The first is a traditional piece of code. Something simple like a bit of marketing text to a page or a new react component that adds another css effect to some content. The second type is a long complex merge request, for something more complex than a menu (not really though)…tabs, uses new dependencies, none of the old dependencies, is filled with emdashed code comments about personal dev choices (instead of logic flow or business context), and the core file convention is named after the implementers library choice: `react-tabs`. If I bring up any of these issues with the implementer they tell me “we can fix it later and they need to just get it out”.

The first type of merge request is one that should be generated by an LLM and the second is one that should be generated by a human.

Instead I get neither but I get efficiency so someone can deliver at the last minute. And so I can can go mop up the work later or my job is hell the next time “we just need to get this out the door”.

THANK YOU LLMS

As the mythical man month says: never ship the prototype. Plan to write one and then throw it away.

From my relevant post last week:

> AI now lets anyone write software, but it has limits. People will call upon software practitioners to fix their AI-generated code.

https://www.slater.dev/about-that-gig-fixing-vibe-code-slop/

In my experience, LLM-generated code is only as good (or as bad) as the software engineering skills of the “vibe coder.” A seasoned engineer will not only craft clear, detailed prompts that specify how something should be implemented, but will also review the AI’s output on the fly, correcting major derailments—things like: “Don’t create a new function for that; just modify X to add support for this case.” They’ll even do an initial review of the code before opening a PR.

The real problem arises when non-technical people use an LLM to generate a full project from scratch. The code may work, but it’s often unmaintainable. These people sometimes believe they’re geniuses and view software engineers as blockers, dismissing their concerns as mere technical “mumbo jumbo.”

That is essentially how I survived the nuclear winter after the dot-com bubble burst (taking out most of the senior level tech workers across the vast majority of US domestic business --in those days, senior level meant formal training, 20+ years experience and over 40yo with kids in school, mortgages, etc..and when all the jobs go away in an industry for six years, you are forced to crack open your 401K and retire while you figure out something else...there was no coming back for them). The takeaway now is that the informally trained web people who came up without guidance beyond Google Search and forum contributions from Europe but gained control of the industry in the Crash's aftermath will live on as "vibe coders" forever and their "heavy lifting" partners at AWS, GCP and Azure will live on for a while hosting trillion+ parameter LLMs, even as the first wave of American CS graduates since the Crash are about to (finally!) hit their 20 year mark and gain control of tech across all industries (pulling the plug on the "heavy lifting cloud" that they don't need/want to budget for).

But unlike that six year gap during the tech nuclear winter (2000-2006) when you could literally follow those over-confident $10/hr kids around cleaning up one botched effort to port custom Windows apps to LAMP after another, this time it will be different. The LLMs are trained largely on the European-dominated code bases on Github and it's just enough to keep the "vibe coders" out of real bad trouble (like porting a financial application from Visual BASIC into PHP which has different precision floating point resolution between distributions/releases or de-normalizing structured customer data and storing it in KV pairs "because everybody is doing it so relational databases must be obsolete".) The work to cleanup their "vibe coded" mess will not be as intense (especially considering LLMs will help), but there will be a lot more of it this time around and re-hosting it more economically will be a Thing.

Sadly, American businesses will discover they don't need trillion parameter LLMs (due to MoE, quantization, agentic mini-models, etc.) and the supply of acceptable vector processing chips will catch up to demand (bringing prices down for "on prem" deployments) and that "AI snake oil factor" (non-deterministic behavior and hallucinations) will become more than a concern expressed over weekend C-suite golf games and yacht excursions (you know, where someone always gets fired to set an example of what happens when you don't make your numbers). AI had been dead so long that the top C-suites can't even remember the details of how/why it died anymore (hint: you could get fired for even saying "AI" up until the 2000 Crash giving rise to the synonym "ML" as a more laser focused application of AI), just that they don't trust it. The astonishing demonstrations at OpenAI, Anthropic, xAI, Google and Meta are enough to cause C-suites to write a few checks, causing a couple of ramps in the stock market, but those projects by and large are NOT working out due to the same 'ole same 'ole and I fear this entire paradigm will suffer the same fate as IBM Watson. The stock market may well crash again because of this horsepucky even though there IS true potential with this technology, just as with Web 1.0. (All it needs for that is a catalyst event --maybe not Bill Gates throwing a chair, maybe something in the dispute between Sammy and Elon.) Same as it ever was.