Trust the AI, says new coding manifesto by Kim and Yegge

DevOps guru and ex-Googler say vibes beat reading diffs but there are risks

"Accept All. Always. Don't read the diffs anymore."

This is OpenAI co-founder Andrej Karpathy's vision of "vibe coding" as he described on X in February – a new approach where developers surrender control to AI agents and simply trust the vibes.

Vibe Coding by Gene Kim and Steve Yegge

Karpathy did not advocate vibe coding for production, describing it as "not too bad for throwaway weekend projects."

Now Gene Kim and Steve Yegge have written the book on it, and they want to persuade coders this isn't madness, but the future of software development.

Kim is a researcher into high-performing technology organizations, well-known for his books on DevOps, and Yegge is a software engineer and blogger, formerly at Amazon and Google, and these days working on AI coding tools at Sourcegraph.

The authors state in the preface that "vibe coding seems to be reinventing the foundations of how we build software." They have set out to convince skeptics by describing their own learning journey, targeting the book at "any developer who is building things" as well as product owners and infrastructure engineers.

The book is presented in four parts. The first, called "Why vibe code," makes the case for this new way to develop software. Advantages cited are not only higher productivity, but the ability to tackle more ambitious projects, experiment more, shift the focus away from implementation details, and reduce the cost of change. In addition, non-technical people can now build software rather than waiting for the developer team to get through its backlog of requests.

Although this section contains more than its share of breathless advocacy, the authors also describe pitfalls. One thing that saves this title from being useless propaganda is that both Kim and Yegge have extensive hands-on experience of and seem open in describing things that went wrong. These issues include:

• That the coding agent had silently deleted or hacked the tests to make them work, and had outright deleted 80 percent of the test cases in one large suite
• That the coding agent generated a giant, 3,000-line function with no modular boundaries, that was impossible to understand or modify
• That the coding agent nearly deleted weeks of work after being instructed to remove unneeded branches from a Git repository

The authors do not see these issues, and others like them, as reasons to abandon vibe coding, but rather as learning experiences that help them to learn how to manage AI tools correctly.

Part two expands on this with a look at the theory and practice of vibe coding, using a somewhat tiresome cooking metaphor involving chefs and sous chefs. Nevertheless there are some thoughtful tips here, such as optimizing AI context and the problem of context saturation: giving AI tools too much context degrades its responses, even to the point of incoherence.

Part three turns to the confusing profusion of vibe coding tools and agents, and how the classic developer loop – code, compile, run, test, debug – changes in vibe coding to one that starts with defining a subtask and continues with an AI conversation, AI-generated plan, and so forth. The importance of tests is repeatedly emphasized.

The final section of the book looks at how organizational culture needs to change to get the best from AI coding. Here Kim is in his element, as he discusses executive strategies, building standards, and changed skill requirements. Communication skills, once a nice-to-have, are now non-negotiable, he writes.

The authors believe that "all knowledge workers will start vibe coding before long," an alarming thought to those who believe coding should be the exclusive province of trained developers.

Despite its tendency to treat vibe coding as a bright new dawn for software, the authors do acknowledge the risks. A statement toward the end of this book says: "We have no doubt that if you adopt vibe coding with reckless abandon, ignoring the practices presented in this book, you are on a surefire path to chaos and endless pager calls – possibly followed by executives being forced to ban vibe coding."

The Reg found the book repetitive in places – despite, apparently, using AI to help manage the content – and the writing style is not for everyone. It is also questionable whether this book is altogether about vibe coding. At one point, the authors describe how the AI got stuck trying to write a Gradle script and Yegge intervened to write the code himself. Most vibe coders are not Steve Yegge.

Despite annoyances, though, there is plenty to learn from the authors about how to do AI-driven coding both individually and as a team or organization. If one takes the view that AI-driven software development is not going away, despite its many risks, then this is a valuable read.

That said, the advocacy that characterizes much of the writing here, especially in its early chapters, is the worst thing about it. A more nuanced approach would have been preferable. Karpathy himself said in a recent interview: "I feel like the industry is making too big of a jump and is trying to pretend that this is amazing, and it's not. It's slop... the models are amazing. They still need a lot of work. For now, autocomplete is my sweet spot."

Vibe Coding by Gene Kim and Steve Yegge is published by IT Revolution, ISBN 9781966280026 (paperback) or 9781966280033 (ebook). ®