Under the name Artcraft, AI-generated copies of well-known software tools are currently being offered – including replicas of Adobe Photoshop and Microsoft Word. The clones can be used free of charge and are not created through classic programming, but largely from prompts, i.e. natural-language instructions to a language model. For companies, this is less a reason to worry about existing licences than a useful benchmark: it shows how robust so-called vibe coding has become for complex applications.

What vibe coding delivers in practice

Vibe coding refers to the development of software in which the generated code is barely read in detail any more, but is judged primarily by the visible result. This works surprisingly well where interfaces and basic functions are required: a text field with formatting options, an image with a few filters, a familiar toolbar – such building blocks are present millions of times in training data and can therefore be reproduced reliably.

This is precisely where the appeal of such clones lies. At first glance they seem familiar because they imitate the operating logic of well-known programs. However, the actual performance of established tools rarely lies in the interface. It lies in colour management, file formats, performance with large documents, undo histories, stability over long sessions and in thousands of special cases that have been incorporated over years. This depth does not emerge from a prompt.

The three realistic limits

For project managers, the limits of prompt-generated software can be narrowed down to three areas.

  • Quality: Functional and production-ready are two different states. Edge cases, error handling, accessibility and behaviour under load are typically what is missing first in generated applications – and at the same time what costs the most in operation.
  • Maintainability: Code that nobody in-house understands is a liability. If an interface needs to be adapted or a security issue closed in a year's time, it takes people who can follow the structure. Generated code without architectural decisions, without tests and without documentation makes this considerably more difficult.
  • Legal certainty: Replicas of established products operate in a delicate area. Trademarks, icons, interface design and file formats are not automatically free to use just because a model can reproduce them. Anyone who wants to use such results commercially should clarify the origin and the licence situation before the application is integrated into a process.

Where AI prototyping pays off

Nevertheless, the productive area of application for this technology is considerable – it simply lies earlier in the project timeline than many expect. Prompt-generated applications are an excellent means of making an idea visible before budget is committed.

  • Concept validation: A clickable prototype answers questions in two days that would otherwise be discussed in two workshops.
  • Internal helper tools: Small tools for a manageable user group, without personal data and without connection to core systems, are a sensible entry point.
  • Requirements clarification: What is noticeably cumbersome in the prototype can be described more precisely in the later requirements specification.
  • Training and coordination: Specialist departments discuss a running interface more concretely than a document.

Where professional development remains necessary

As soon as an application is used productively, the requirements shift. Then it is about data protection, access rights, logging, backups, update capability and the question of who is liable when something fails. These topics are not add-ons, but the actual effort of professional software development.

This applies in particular to systems that dock onto existing infrastructure: a TYPO3 project with grown content structures, an ERP system, a CRM. Here it is not the interface that decides success, but the clean interaction with what is already there. A generated draft can shorten the path there – it cannot replace it.

A pragmatic approach

A clear separation of two phases makes sense. In the exploration phase, AI may be used generously: fast, cheap, disposable. What emerges there is deliberately treated as throwaway work and not silently carried over into productive operation.

In the implementation phase, AI remains a tool, but under human control: with code reviews, tests, defined architecture and documented decisions. The speed gains are real there too – they simply arise differently, namely through assistance rather than autonomy.

A useful touchstone before any decision: who will take responsibility for this software in three years? If the question cannot be answered, the prototype is not yet a product.

Classification for web and software projects

AI-generated clones of well-known tools show above all how cheap interfaces have become – and how valuable everything else therefore becomes: data model, integration, operation and legal clarity. For web and software projects, this shifts the focus from the question of how something looks to the question of how it works in the long term.

Anyone who uses AI early on for experimentation and then has the load-bearing parts built cleanly gets the best of both worlds.

Sources