Under the name World of Claudecraft, a project has emerged that serves as a vivid example in discussions about AI-assisted development: a so-called vibe coder — that is, someone who creates software primarily through instructions in natural language to an AI model instead of writing every line themselves — has built their own version of the online role-playing game World of Warcraft with it. The effort: two days and a few hundred US dollars, specifically around 200 dollars for using the model.

The result is playable. According to the available reports, however, it does not come close to the original, which was developed over years by a large studio and has been operated and further developed for decades. The value of this example lies precisely in this area of tension: it shows both — the new speed at which software comes into being and the point at which speed alone no longer helps.

Why this is relevant for companies

A game clone sounds far removed from customer portals, booking systems or internal tools at first. The comparison is still worthwhile because an online role-playing game is technically no trivial task: it requires three-dimensional rendering, controls, game rules, state management and server logic. If something like this can be brought to a runnable state in days, that shifts expectations about how long a first working proof of concept in a project may take.

For decisions, this means above all one thing: the cost of trying out an idea has dropped significantly. Where a concept paper and a click dummy — a non-functional mock-up of the interface — once marked the state of discussion, today a real, usable prototype can often be available. Discussions about usability, the data model and missing requirements become more concrete as a result and take place earlier.

What such a prototype achieves

  • Demonstrate feasibility: A runnable state answers the question of whether an idea fundamentally works more reliably than any presentation.
  • Sharpen requirements: Once people can touch something, the comments come that nobody formulated in the workshop.
  • Compare variants: If a draft is inexpensive, you can build two or three instead of committing to one in advance.
  • Justify budgets: A visible interim result makes investment decisions comprehensible for committees.

And what it does not achieve

The decisive difference between a prototype and a product lies not in the features you see, but in what lies beneath them. A two-day-old state typically has no robust rights and role logic, no reviewed data protection architecture, no load tests, no test coverage, no documentation and no answer to the question of who will maintain it in three years. The fact that the example project falls short of the original in several respects is therefore not a side note, but the actual finding.

There is also an effect that is regularly underestimated in practice: prompt-driven development quickly produces a lot of code. Whether this code is consistently structured, whether dependencies are chosen cleanly, whether security vulnerabilities are contained and whether the whole thing can be extended later is not decided by the prompt, but by the assessment of the result. This assessment requires experience. Those who lack it cannot recognise the difference between "works" and "sustainable" — and only notice it when changes become expensive.

A pragmatic approach to the tool

A clear separation between two phases makes sense. In the exploration phase, AI-assisted development may go very far: here speed counts, not elegance, and throwaway code is expressly permitted. The only important thing is that everyone involved knows that it is throwaway code.

In the implementation phase, the standards change. Architecture decisions, data models, interfaces, permissions, the operating concept and testability belong in the hands of people who are responsible for the result. AI remains useful there too — when writing tests, when refactoring, for routine tasks, for understanding unfamiliar code. Only responsibility for the structure cannot be delegated.

For project planning, this leads to an uncomfortable but honest consequence: if the prototype is finished in days, the majority of the effort shifts to the later stages. It does not disappear. What makes a project expensive is rarely the visible features, but integration into existing systems, migration of data, accessibility, operations and maintenance over years.

Assessment

For web and software projects, this means: use AI-assisted prototyping early and proactively to test ideas before budgets are committed — the hurdle for this is lower than ever. But do not expect the time from prototype to a productive, maintainable and secure system to shorten to the same extent. The gain lies in better decisions at an earlier point in time, not in the elimination of development work.

Sources