Spec‑Driven Development – How to Stop AI from Rewriting Your Codebase

AI-assisted coding is powerful — and sometimes a little too powerful. You ask an LLM to add a small feature, and suddenly half your codebase looks different. Sound familiar?
In this talk, Kirill walks through a real example from one of his own projects: a simple request to “add one button” that unexpectedly triggered changes across code, architecture, and even product behavior. Using this case study, he shows why prompt‑only workflows often lead to unpredictable results.
The second part introduces a spec‑driven workflow as a practical way to regain control. Instead of prompting the model directly, the task is defined through a clear spec with scope, constraints, and acceptance criteria. The same feature is implemented twice — once with a simple prompt, once with a spec — and the difference is striking.
This session is hands‑on, based on real code rather than demos, and ideal for developers who want to use AI productively without losing control of their codebase. Talk duration: ~30 minutes plus discussion and experience exchange.

BIO:
Kirill is a backend engineer based in Munich, working with .NET and distributed systems.

Recently he has been focusing on AI-assisted development in real-world projects, exploring how to use LLMs without losing control over code quality and architecture.
He shares practical experience from his own projects, focusing on what actually works in production rather than theory or demos.