Documenting software architecture reader-friendly with LLMs
AI significantly simplifies documentation for software developers and architects. It writes stories, creates graphics, and revises its output.
(Image: KorArkaR / Shutterstock.com)
In his presentation “AI-centric software documentation: Insights into the practice of AI startups” at the online conference betterCode() ArchDoc 2025, Ingo Eichhorst presented concepts for the automated documentation of architecture projects. He shows practical examples from his work in AI startups.
The approaches he presented range from automated quality assurance through simulated developer interactions to self-documenting systems. Modern documentation systems can not only capture knowledge but also increase the resilience of the entire software system.
His approach starts with the specification of a project, which not only describes the tasks the software needs to perform but also provides additional required data from the context, such as architectural guidelines, domain models, or company specifics: “Give the model the information you yourself would need to complete the task,” is the speaker's motto. The specification becomes the single source of truth, and developers subsequently work only with this.
(Image:Â NeuralStudio / Adobe Stock)
The online conference betterCode() ArchDoc by iX and dpunkt.verlag on May 20, 2026, will present lightweight documentation concepts such as the arc42 Canvas or Docs-as-Code for working like programming.
AI for automating documentation will also be a central theme of the conference again.
The concrete implementation of the code and the associated documentation, however, is left to the LLM. Concrete tools for AI can include a repo map and an arc42 template. Changes required by correction loops also only take place in the specification. The decisions for these architectural adjustments result in more or less complete Architecture Decision Records.
Graphics and Storytelling
With the help of AI, there are many possibilities for the concrete implementation of documentation: the LLM takes over a target group-oriented reader perspective and writes reader-friendly, easily accessible stories about the project, thus reducing cognitive load. The associated graphics can also be generated automatically, for example, from JSON data.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier eine Vimeo-Video (Vimeo LLC) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Vimeo LLC) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
Even an evaluation of the documentation can be done by AI itself, for example, through digital twins of people. The LLM then automatically corrects its work upon request. At the end of the presentation, Eichhorst presents a project in which the AI revised a complete documentation. This allows a proliferation of branched information distributed across many web pages to be systematically consolidated.
(who)
