Ungrounded AI defaults to the ordinary blend of its training data. This book grounds it in the published work of the field's experts, and keeps it there across a whole build.
Ask a model to build something and give it nothing else, and you get the blend of everything it saw in training: far more of it ordinary than expert. It will compile. It will pass casual inspection. And it will sit at a ceiling you can't see, because the output reads the same whether it reflects the field's best practice or merely sounds competent.
The Grounded Vibe Methodology begins with an effect the author noticed, tested deliberately, then traced how far it reached: name the experts whose published work covers what you're building, and the model's output shifts toward that work. Say “follow Tufte and Few” and the information displays begin to track the principles those authorities set down. Name Fowler on refactoring, Beck on test discipline, Wiegers on requirements, and the output moves toward the documented practice of the field.
What the book adds
That much is one sentence in a prompt. It is also where most people stop, and it isn't enough. Naming raises the discipline of a single reply; it does nothing to keep that discipline holding across requirements, design, build, and review, and nothing to stop the model building the wrong thing well. The book is the structure that does:
- The two failures every AI build hides: building the wrong thing, and building the right thing carelessly. Each passes casual review for a different reason.
- Why the obvious fixes fall short. Personas (“you are a senior architect”), spec-driven development, retrieval, agentic test loops. Each helps, and each leaves something untouched.
- A fourteen-phase pipeline that grounds both what the model generates and how its output is reviewed: expert panels, defect-class partitioning, mechanical gates outside the model's discretion, and a release verdict.
- The evidence. In a controlled experiment across three model families, ungrounded AI review ranked expert-designed work backwards or could not tell it apart; review grounded in named principles saw the difference clearly. Two worked case studies show real software built with the methodology, not just an experiment run on it.
Casual inspection has no standard to measure against, so it approves work an expert-informed review would reject, and cannot detect the gap from inside its own frame. Grounding supplies the standard.
The name is a deliberate riff on vibe coding.
Who it's for
Written for three readers: practitioners who want to raise the standard of what their AI tools hand back; engineering leaders weighing adoption across a team; and researchers, who will find the effect documented with its limits stated, including what remains unexplained.
You do not need to know which experts to name in advance: a discovery protocol surveys your domain and stack and proposes a scored roster for your confirmation. And you do not have to commit to the whole method to begin: it ships as commands you run inside an AI coding assistant, and each one (review a document, audit a codebase, sharpen requirements, draft a spec) works on its own. Point one at a document or a codebase and see the standard shift. Keep the vibes. Ground them in the canon.