Developer Felix Rieseberg has described an AI-directed production workflow that used roughly 60 cloud sessions to redesign his personal website, create a detailed three-dimensional room and synthesize supporting media. The project was managed through a Claude Project, with much of Rieseberg’s review and feedback taking place on his phone rather than through locally running development tools.
Rieseberg says the project began with collections of visual references, including posters, films, advertisements and personal photographs. Ten sessions explored alternative directions in parallel. When early concepts felt too derivative, he asked for further variations and eventually combined an AI-suggested VHS-tape gallery with his own idea of a room inspired by a 1990s German film setting.
The second draft arrived 39 minutes after the first brief and was already recognizably related to the version that eventually shipped, according to his account. Reaching the final result took about 60 iterations. Each iteration could be published as a private preview page, allowing Rieseberg to inspect alternatives and send criticism from either a computer or phone while the underlying work continued in cloud-hosted sessions.
Blender became central even though Rieseberg says he does not know how to operate its graphical interface. Instead, Claude generated geometry and controlled Blender through Python. The completed room is represented by around 3,000 lines of Python, with separate components structured so multiple sessions could work on elements concurrently. Some furniture assets came from the CC0-licensed Poly Haven library, while the desk, televisions, music players, old computer and room geometry were modeled within the workflow.
The visual system also required more than simple object generation. Procedural textures represented wood, plaster, plastic and metal. To achieve softer lighting in a browser, the workflow used Blender’s Cycles renderer to bake lighting into texture maps. Static surfaces received additional UV layouts and several maps representing different light sources and states. Moving objects and screens remained dynamically lit.
Rieseberg frames the project as an example of shifting from direct tool operation toward directing goals, reviewing options and supervising specialized sessions. He used Opus 5.5 and says individual threads could install tools such as Blender, FFmpeg and Playwright, share files and memory, check pages with browser screenshots, and publish previews.
The account is a first-person project report, not a controlled productivity study, and it does not establish that the approach will generalize to other teams or workloads. It does show how cloud execution, shared project context and parallel agents can extend AI-assisted development beyond source-code editing into 3D production, sound and iterative visual design.
