LLM Coding with Semantic Anchors: From Vibe Coding to a Real Java App with Ralf D. Müller

Most LLM-based apps struggle because they rely on vague prompts and “vibe coding.” In this live coding session, we show how to build a *real Java application* using *Semantic Contracts and Semantic Anchors* to make LLM-driven systems *more predictable, structured, and reproducible*.

Written summary: https://rabauer.dev/en/blog/semantic-anchors/

In this session, I’m joined by @ralfdmueller, committer at arc42 and creator of docToolchain, to explore a more structured approach to building applications with GenAI and LLMs.

Instead of treating LLMs as unpredictable black boxes, we introduce *Semantic Anchors* and *Semantic Contracts* to define clear expectations and guide behavior with well-defined structure. This helps move from trial-and-error prompting toward systems that are easier to understand, test, and evolve.

### 🧠 What you’ll learn

* Why “vibe coding” breaks down in real-world applications
* How *Semantic Anchors* help guide LLM behavior
* How *Semantic Contracts* improve consistency and clarity
* How to structure LLM interactions in a more engineering-driven way
* How to apply basic software architecture thinking to GenAI apps

### What we build

We implement a simple *train scheduling application* for model railroads where users can:

* Define trains and stations
* Create and manage schedules
* Interact with the system using structured LLM workflows

The technical setup includes:

* @java + @Quarkusio backend
* @vaadinofficial UI
* PostgreSQL database

### Why this matters

Many GenAI tutorials focus on prompt tricks. This session focuses on *building systems you can reason about*:

* Clear contracts instead of hidden assumptions
* More reproducibility instead of pure randomness
* Structure instead of guesswork

This is especially useful if you’re starting to apply *software architecture principles to LLM-based applications*.

00:00 Introduction
06:22 Semantic Anchors
17:40 Spec-Driven Development
21:12 Add Feature Request - TDD Hamburg School
24:09 Semantic Contracts
27:20 Start Coding with Semantic Contracts and Docker sbx
34:22 Creating Requirements with the Socratic Method
51:20 Vibe-Coding Risk Radar
01:04:00 Installing docToolChain (small failure ^^)
01:11:05 Create an arc42 document from the Requirements
01:26:00 Looking at one ADR in depth
01:29:00 Core advantage of Semantic Contracts
01:32:15 Create the Specification
01:42:40 Create Epic- and Story-Issues for implementation
01:43:51 Why use CLAUDE.md for Semantic Contracts and not Skills?
01:45:26 What about Stale Documentation?
01:48:45 Cucumber or Gherkin tests
01:49:20 Do LLMs ignore tests?
01:50:58 Check the Issues
01:54:05 Analyse Issues through the LLM
02:00:00 Implementing the Issues
02:15:20 Summary / Conclusion

Видео LLM Coding with Semantic Anchors: From Vibe Coding to a Real Java App with Ralf D. Müller канала Johannes Rabauer