Website of Peter Westerman

The term “vibe coding” was coined by AI researcher Andrej Karpathy in early 2025. It describes a style of software development where you describe what you want to build in plain language, an AI writes the code, and you iterate — testing, evaluating, directing, correcting — until the software does what you need. You do not need to understand every line of code the AI produces. You need to understand what you are trying to build and whether what you are getting back is working.

The name captures something real about the experience. You are not writing code. You are not managing a developer. You are communicating intent, evaluating output, redirecting when the result is off, and iterating until it converges on what you need. It feels more like directing a creative collaborator than either traditional programming or traditional management. The “vibe” is the intent — the purpose, the feel, the outcome you are aiming at — more than the technical specification of how to achieve it.

But the most important thing about vibe coding is not what it is in a software context. It is that the core skill — communicating intent precisely, evaluating output rigorously, iterating systematically — is not specific to software at all. It is the core skill of working effectively with AI on any task. The people who develop this skill in a coding context have learned something that transfers immediately to research, writing, analysis, strategy, legal work, and every other domain where AI can produce useful output.

What a Session Actually Looks Like

A session starts with a requirements document — a clear description of what needs to be built, who will use it, and what success looks like. Not a formal specification. A structured description with specific outcomes, user scenarios, technical constraints, and an explicit list of what is in scope for this build versus what comes later. The quality of this document is the primary predictor of first-pass output quality.

That document goes to the AI — in our case, Claude running in the Cursor IDE — and the AI proposes an architecture: what files to create, how data flows between them, what the database structure should look like, how the user interface should behave. The human reviews the proposal and either approves it, modifies specific elements, or rejects it and redirects toward a different approach. This architecture conversation is not overhead — it is where the most expensive mistakes get caught before they are built in.

Then the AI builds. It writes code across multiple files simultaneously, implementing the architecture it proposed. The human watches, catches things that seem structurally wrong, waits for the build to complete, tests the output in a real environment, and reports what is working and what is not. That report goes back to the AI, which addresses the issues. That loop — build, test, report, fix — runs multiple times per session and multiple times per day.

What the human is doing throughout: judgment. Deciding what to build. Evaluating whether what was built is right. Catching errors in logic or design that the AI’s technical correctness cannot prevent. Redirecting when something is going in a direction that will not work. The code generation is handled by the AI and is nearly instantaneous. The judgment is irreplaceable and is where all the human time goes.

The Tool Stack That Enables It

Effective vibe coding requires a tool that lets the AI see and modify your entire project simultaneously, not just respond to questions one at a time. The difference between a general-purpose AI chatbot and a purpose-built AI coding environment is significant.

We used Cursor — a code editor built specifically for AI-assisted development. Cursor lets the AI see every file in the project, understand the relationships between them, and make coordinated changes across multiple files in a single response. When a database schema changes, the AI understands which other files reference that schema and updates them consistently. This is architecturally important: it means the AI is working on your project, not on a pasted excerpt from your project.

The AI model was Claude — Anthropic’s large language model. The choice of model matters less than the capability profile: you need a model that maintains context across long, complex conversations, reasons about multi-file systems coherently, and applies consistent conventions once they are established. The third tool is version control — GitHub repositories for every product. Version control provides a safety net for changes, a history of decisions, and a deployment pathway. It is not optional in any serious practice.

The Skill Gap That Most People Miss

The biggest misconception about vibe coding — and about AI collaboration generally — is that the hard part is getting the AI to produce good output. It is not. Current AI systems produce high-quality output on well-defined tasks consistently. The hard part is defining the task well enough for the AI to do it, and evaluating the output rigorously enough to know whether it is right.

These are communication skills, domain expertise, and judgment — skills that experienced professionals in almost any field already have. A lawyer who can articulate exactly what a contract clause needs to accomplish and then evaluate whether a draft clause does that is already equipped to direct AI for contract drafting. A financial analyst who can specify exactly what a model should calculate and then audit the output against expected behavior is already equipped to direct AI for financial modeling. The domain expertise transfers directly. What needs to be learned is the communication interface — how to structure intent for an AI collaborator, how to evaluate AI output systematically, and how to build the context infrastructure that makes the collaboration compound over time.

The context infrastructure piece is what separates professionals who get marginal improvement from AI from those who get transformational leverage. Building and maintaining context files, developing template libraries, creating specialist AI profiles for recurring task types — these are the practices that produce compounding returns. Without them, AI collaboration produces decent one-off results. With them, it produces a systematically higher-leverage practice that improves with every session.

Why “Vibe” Is the Right Word

The “vibe” in vibe coding captures the nature of the communication shift. Traditional software development required translating intent into precise technical specification before any work could begin: detailed functional requirements, API contracts, database schemas. All of the translation from “what I want” to “how to build it” happened in the requirement-writing phase. The developer executed the specification.

In vibe coding, most of that translation happens in real time during the collaboration. You describe the vibe — the purpose, the user experience, the outcome — and the AI handles the translation to implementation detail. You redirect when the translation goes wrong. The human’s job is to hold the intent clearly and evaluate whether the output honors it, not to do the translation themselves.

That shift — from detailed specification to intent communication with real-time evaluation — is also the shift happening across knowledge work broadly. The professionals who thrive will be those who become excellent at holding and communicating intent clearly, and at evaluating whether AI-produced output honors that intent. The ones who struggle will be those whose primary professional value was in the translation layer: the production work that AI now handles.

The Non-Software Version

Every software-specific lesson in this series has a non-software equivalent, and we will draw those equivalents explicitly throughout. Vibe coding for a marketing strategy: the marketer describes the campaign’s audience, goal, competitive context, and success criteria; the AI produces a strategic framework; the marketer evaluates it against their domain knowledge of the brand and audience; they iterate. Vibe coding for financial analysis: the analyst specifies the analytical question and the modeling approach; the AI builds the model; the analyst evaluates the output against their expectation of what the answer should be given the underlying business; they refine.

The five-phase methodology — define specifically, contextualize thoroughly, direct precisely, evaluate rigorously, iterate systematically — is the same in every domain. The domain expertise that makes evaluation possible is yours. The production capacity that makes iteration fast is the AI’s. The combination produces leverage that neither can achieve independently.

This series documents that combination applied to software. Extract the methodology. Apply it where you work. The leverage scales with how clearly you can define what you need and how rigorously you can evaluate what you get. Both of those capabilities improve with practice, and they improve faster when the feedback loop is as compressed as vibe coding makes it.

The Difference Between a Chatbot and a Coding Partner

One of the most common misconceptions about vibe coding is that it is essentially the same as using ChatGPT to generate code snippets — paste in a problem, get back a solution, copy the solution into your project. That is not vibe coding. It is one-shot code generation, and it is substantially less powerful because it lacks the project context, the multi-file awareness, and the session continuity that make vibe coding effective for building real products.

Cursor, the IDE used throughout this case study, changes the nature of the interaction fundamentally. The AI can see every file in your project at once. It understands which files import which other files. It knows what functions exist and what they do. It can make coordinated changes across ten files in a single response, keeping everything consistent. When you change an interface definition, it updates all the implementations. When you rename a database column, it updates all the queries that reference it.

This multi-file, project-aware context is what makes building real products possible rather than just generating code fragments. It is also what makes the context file so important: the AI is not just generating isolated code, it is making decisions about a coherent codebase, and the quality of those decisions depends on how accurately it understands the codebase’s structure and conventions.

The Role of Iteration in Quality

A critical aspect of vibe coding that distinguishes it from one-shot code generation is the centrality of iteration. No single AI response produces production-quality output for a complex feature. The first response produces something close. The second iteration refines it. The third might catch an edge case. The fourth might improve the error handling. The product that ships is the result of that iteration process, not the first response.

This means that the human’s role in iteration — testing, evaluating, providing specific feedback, directing corrections — is not a workaround for AI limitations. It is the methodology itself. The AI provides the generation capacity that makes fast iteration possible. The human provides the judgment that makes iteration converge on the right answer. Neither can achieve the same result alone. The methodology is the combination.

Understanding this changes how you think about sessions. A session that produces a complete, working feature in four iterations over ninety minutes is not a session where the AI was “good enough to get it right on the fourth try.” It is a session where the methodology worked as intended: rapid generation followed by rigorous evaluation followed by targeted correction, repeated until convergence. The iteration is the process. Speed comes from the generation being fast. Quality comes from the evaluation being rigorous.

Why Domain Experts Have a Natural Advantage

In a vibe coding practice, the human’s domain expertise is what determines whether the AI’s output is correctly evaluated. A developer who does not understand security cannot evaluate whether an AI-generated authentication implementation is secure. A marketer who does not understand audience psychology cannot evaluate whether an AI-generated campaign brief will resonate. A financial analyst who does not understand the underlying business cannot evaluate whether an AI-generated model’s outputs are plausible.

This is why vibe coding — and AI collaboration generally — produces better outcomes for domain experts than for novices. The novice gets plausible output that may be wrong in ways they cannot detect. The expert gets better-evaluated output that converges toward correct through their informed correction cycles. The AI’s capability is the same in both cases. The human’s evaluation capability determines how much of that capability is extractable.

This also means that the best investment in improving AI collaboration outcomes is improving domain expertise, not just AI communication skills. The professional who deepens their understanding of their domain becomes better at evaluating AI output in that domain. That improved evaluation capability directly improves the quality of the final outputs produced through AI collaboration, because the iteration cycles converge more accurately on what is actually right rather than what is merely plausible.

How Recent AI Innovations Change This Picture

The definition of vibe coding described in this post — conversational, intent-driven, iterative collaboration — remains accurate. What has changed is the tooling available to implement that collaboration and the scale at which it can operate.

Agent Skills, introduced by Anthropic in October 2025, are reusable folders containing instructions, scripts, and resources that can be loaded into any Claude session. This is the formalized, platform-supported version of what this case study was building manually with CLAUDE.md context files and shared library documentation. The difference is portability and composability: skills can be shared across projects, across teams, and across different Claude applications without manual copying and maintenance. What took significant infrastructure effort to implement manually is now a built-in feature of the platform.

The Model Context Protocol (MCP), Anthropic’s open standard for connecting AI applications to external systems, changes what a vibe coding session can reach. Instead of pasting content from external systems into your AI chat manually, MCP lets Claude connect directly to your database, your project management system, your version control, your analytics platform. The AI’s awareness of your project context expands from the files on your local machine to the live systems your work depends on.

Extended thinking — Claude’s ability to reason step-by-step through complex problems before generating output — changes the quality ceiling on architectural decisions. The architecture conversations described in this post would benefit substantially from extended thinking: a Claude instance reasoning through the tradeoffs of different database schemas, API structures, or component hierarchies before committing to a recommendation. The human review step becomes even more valuable when the AI’s proposed architecture is the result of deep deliberation rather than pattern matching.