Blog

Gedanken zu AI, Software-Architektur, Dokumentation und der Zukunft der Softwareentwicklung. Ursprünglich auf LinkedIn veröffentlicht.

05 Jul 2026
ENLLMGenAI

Your LLM Holds a Whole Library. You Are Shown One Shelf.

An LLM understands hundreds of languages equally well, yet on your question it hands you back only the world that matches your own language. A five-language experiment shows the bias sits in the raw generation, not only in search, and depends on the topic.

Artikel lesen
21 Jun 2026
ENAgentic CodingdocToolchainOpen SourceGenAI

Three Days. I Did Not Type a Single Line of Code.

Three days. 158 commits. 44 pull requests merged. I did not type a single line of code. What a weekend on docToolchain main-4.x looked like when the bottleneck stopped being typing — and what it means for unpaid, single-maintainer open source.

Artikel lesen
10 Jun 2026
ENGenAI

Same Prompt, New Image Model: Switching to gpt-image-2

Same prompt, same references, same character description. The only variable was the image model. Top row gpt-image-1.5, bottom row gpt-image-2. Character consistency caught up, composition got more deliberate, and I switched the default in my image-generation skill.

Artikel lesen
10 Jun 2026
ENSemantic AnchorsLLM

An Anchor Delivers Only as Far as the Prior Reaches

A pull request asked why our anchor still points at Cockburn instead of newer Use-Case 2.0 and 3.0. We turned it into an experiment. The result was Claude Haiku confidently printing a "Use-Case 3.0" document whose body was pure Cockburn from 2001. No error, no hedge. This is the failure mode no one notices.

Artikel lesen
26 Mai 2026
ENSemantic AnchorsAgentic CodingLLMSoftware Architecture

Trust Is Good, Verification Is Better

Most agentic coding rests on trust. The harness is what turns "it ran" into "I will deploy this". This is the full inventory of layers, what each catches, and what no harness can.

Artikel lesen
18 Mai 2026
ENSemantic AnchorsLLMarc42Software Architecture

My AI Improved Its Own Process

An LLM reconstructed the docs of a 43,000-line codebase. When the verification failed, the fix was not a better prompt. The AI rewrote the rules it works by. A correction loop, not a prompt.

Artikel lesen
28 Apr 2026
ENdocToolchainSoftware ArchitectureGenAIOpen Source

Docs-as-Code Is Now AI Infrastructure

Your LLM assistant starts every session like a new colleague on their first day. Docs-as-Code is no longer a convenience. It is infrastructure. Bausteinsicht is now open source.

Artikel lesen
19 Apr 2026
DESoftware Architecture

Agentic Coding vs. CNC Fräsen

Panel-Diskussion bei betterCode Modern Architecture 2026: CNC-Analogie, Senior/Junior-Produktivität mit KI und autonome Agenten — mit Ingo Eichhorst und Ferdinand Adé.

Artikel lesen
11 Apr 2026
ENDigital IndependenceGDPROpen Source

I DID it!

Replacing Zoom with a European alternative. A systematic Pugh matrix comparison of 7 video conferencing tools.

Artikel lesen
23 Mär 2026
ENSoftware ArchitectureDocumentation

arc42: My Cockpit for AI-Generated Code

We spent 30 years begging developers to write documentation. Now AI writes it for free. And we're still staring at the code. arc42 as verification instrument for AI-generated code.

Artikel lesen
10 Mär 2026
ENAgentic CodingOpinion

The IDE Is Dead

Die klassische IDE verliert ihre Relevanz. AI-Agenten brauchen keine GUI, sie brauchen CLI-Tools und Feedback-Loops.

Artikel lesen
04 Feb 2025
DEOpinion

Meine alternative Sicht auf KI Modelle

Die Medien haben eine verschobene Sicht: Claude Sonnet übertrifft OpenAI für Software-Entwicklung. Mistral und Black Forest Labs führen Leaderboards an. Europa hat kompetitive KI-Fähigkeiten.

Artikel lesen
04 Dez 2024
DEAgentic Coding

Claude Code auf dem Raspberry Pi

Claude Code auf dem Raspberry Pi installiert, Netzwerk gescannt, Shelly-Geräte gefunden. Dann eine Debatte über Dämmerungszeiten - Claude hat gewonnen. Ich schulde Claude ein Bier.

Artikel lesen