#ai-native

Interne Doku veraltete, sobald sie fertig war. Jetzt schreibt und pflegt KI sie und nutzt sie als Kontext. Aus der ältesten Pflicht wird ein Kreislauf.

AI-native ist kein Marketing-Etikett, sondern eine andere Art, Software zu bauen, zu betreiben und auszuliefern. So sieht das in der Praxis aus.

Google's AI design tool generates high-fidelity UI from prompts, sketches, and voice. The interesting part is DESIGN.md and its MCP bridge to your code.

Paper eliminates the design-to-code translation by making the canvas itself HTML and CSS. What you design is what ships. The question is whether that's always what you want.

Pencil puts a Figma-like canvas inside your IDE with design files in Git and AI agents reading specs through MCP. The design never leaves the codebase.

Three new tools close the gap between design and code from different directions. The deeper shift: agents need a shared design vocabulary to work at all.

Everyone obsesses over models and prompts. The real bottleneck in RAG is retrieval, and fixing it takes engineering, not bigger context windows.

A plain-language introduction to Retrieval-Augmented Generation: what it does, how it works, and why it is becoming the default way to use AI on your data.

Andrej Karpathy's autoresearch lets AI agents run about a hundred ML experiments overnight on one GPU. It looks like the beginning of autonomous science.

Both are CNCF Graduated, and both do GitOps well. Their architectures disagree about who the primary operator is. That matters once agents enter the loop.

Most architecture documentation is either absent or abandoned. Arc42 gives it structure that humans can navigate and AI agents can maintain. Here is how we use it and why it stuck.