Pilot-Durchführung: Date to follow · Location to follow · CHF 200 instead of regular CHF 1,200

1 day On-site Max. 12 Teilnehmer

Spec Driven Design with AI

How senior devs use AI coding assistants productively — without hallucinations, regressions and corrupted tests.

Reserve pilot seat (CHF 200) Request in-house training

What this course is not

Not an introduction to AI coding assistants — no tool demo, no 'write your first function with Copilot'
Not an overview of available AI tools — we assume you have already moved past that phase
Not for developers without active experience with at least one AI coding assistant

1 Für wen

Senior developers who already actively use AI coding assistants
Tech leads and engineering managers who want to establish AI-assisted development in their team
Teams using AI productively — and struggling with hallucinations, regressions and corrupted tests

2 What you can do after the course

Introduce a reliable spec-first workflow in your team — instead of waiting for the next vibe-coding miracle
Validate AI-generated code automatically against the spec — with reviewer agents, not with hope
Assess which parts of your existing codebase are suitable for agentic development — and which are not
Slice architecture so AI agents can work reliably — instead of getting stuck in monolithic complexity
Articulate the difference between a productive senior dev with AI and a helpless prompt-typist — and fix the latter in your team

3 Content in detail

Understand why vibe coding fails structurally (stochastic, stateless)
Use specs as single source of truth: Markdown, Mermaid, interfaces, BDD
Apply business analysis formats for LLMs: state machines, decision tables, invariants
Master prompt, context, intent and spec engineering
Understand AI-ready architecture: SDR principles, vertical slices, SCS
Set up reviewer agents: validate AI code automatically against specs
Derive green-field and brown-field strategies for AI-native engineering

Agenda

Morning: Foundations & first spec

Why vibe coding fails: hallucinations, regressions, corrupted tests
AI coding workflows: inline, agentic, spec-driven, dark factory
SDD principle: spec as single source of truth, separating intent from implementation
Business analysis for LLMs: Mermaid state machines, decision tables, BDD, invariants
Exercise 1: Prompt a complex feature vaguely — feel the pain point
Exercise 2: Turn the same feature into a full spec with data models, state machine and definition of done

Afternoon: Agentic engineering, architecture & strategy

Four disciplines: prompt, context, intent and spec engineering
AI architecture constraints: context limits, SDR principles, vertical slices, SCS
Green-field vs. brown-field: strangler fig, discovery before mutation, zero-trust generation
Exercise 3: Spec + agent → plan before code → review with reviewer agent
Exercise 4: Refactor code from Exercise 3 according to SDR principles
Exercise 5: Dark Factory — hand spec to a fresh agent, let it implement autonomously

Method

Five exercise blocks build on each other: the feature from Exercise 1 runs through to the Dark Factory simulation in Exercise 5. All exercises can be done with your own project. Participants without their own code will be provided with an exercise project.

Prerequisites

Several years of professional software development. Active experience with at least one AI coding assistant (Copilot, Cursor, Claude Code, ChatGPT for code, etc.). You should be able to read and evaluate TypeScript or Python code. We do not cover basics of Git, testing or architecture.

FAQ

What tool should I bring? Nothing — we work in GitHub Codespaces. All you need is a browser and a GitHub account.

What if I can’t bring my own code? An exercise project will be provided. All exercises are fully runnable with it.

Is this relevant for brown-field and legacy codebases too? Yes. The afternoon explicitly covers brown-field strategies — strangler fig, discovery before mutation, zero-trust generation for existing codebases.

How does this differ from a Copilot workshop? A Copilot workshop shows you how to use AI tools. This course shows you why that alone is not enough — and what specification and architecture discipline is required for AI code to work reliably in production codebases.

What if the pilot run doesn’t happen? Full refund. No questions asked.

When does the regular course take place? The regular run is planned for Q3/Q4 2026. Pilot participants get priority registration.