DDC 2025: p"AI"r Programming statt Vibe-Coding

Slides for my session at DDC 2025 in Cologne

Transcript

1. .NET Developers Conference 2025 p”AI”r Programming statt Vibe-Coding Wie hole ich das Beste aus AI-Assistenten heraus? Sebastian Gingter Developer Consultant – Thinktecture AG sebastian.gingter@thinktecture.com
2. ▪ What to EXPECT ▪ Overview of coding assistance powered by Large Language Models (LLMs) ▪ Pragmatic use cases ▪ Suggestions on how to leverage AI tooling for the best ▪ What not NOT TO EXPECT ▪ Claude / Codex / Copilot deep dives ▪ ML & AI fundamentals p”AI”r Programming statt Vibe-Coding: Wie hole ich das Beste aus AI-Assistenten heraus? Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
3. p”AI”r Programming statt Vibe-Coding Wie hole ich das Beste aus AI-Assistenten heraus? Sebastian Gingter Developer Consultant sebastian.gingter@thinktecture.com .NET Developers Conference 2025
4. ▪ Generative AI in business settings ▪ AI driven Developer Productivity & Software Quality ▪ All things .NET ▪ Microsoft MVP for .NET & Azure AI Services ▪ sebastian.gingter@thinktecture.com ▪ https://www.thinktecture.com Sebastian Gingter Developer Consultant @ Thinktecture AG Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
5. Agenda ▪ AI Assisted Coding ▪ Why does it work & when not? ▪ What are our tools? ▪ Behind the scenes ▪ Expectation Management ▪ Pair Programming Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
6. Agenda ▪ Potential problems ▪ Possible solutions ▪ Context is king ▪ Conclusion Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
7. AI Assisted Coding Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
8. Why does it work? ▪ Language Models are great at languages ▪ Programming languages: very simple syntax & grammar ▪ LLMs usually are trained to follow instructions very well Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
9. When does it not work? ▪ Being cheap. ▪ Good reasoning and good code output requires using good (= large / expensive / non-local) models ▪ You want (or need) Sonnet / Opus 4.5, GPT-5.1 (Codex 5.1), Gemini 3 Pro (Preview)… Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
10. Different Tools ▪ IDEs ▪ IDE Plugins ▪ CLIs ▪ “Tastes” ▪ Subscription-Based ▪ Bring your own Key (ByoK) ▪ Cloud and local LLMs Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
11. IDEs ▪ Visual Studio 2026 /w Copilot ▪ VS Code /w Copilot (Microsoft / GitHub) ▪ Cursor ▪ Windsurf (previous Codeium) ▪ Kiro (AWS) ▪ Void (Open-Source Cursor Alternative) ▪ JetBrains (all) Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
12. IDE Plugins ▪ GitHub Copilot (VS, JetBrains, VS Code) ▪ Kilo Code (VS Code, JetBrains, Cursor) ▪ Cline (previous Claude Code) (VS Code) ▪ Tabby (VS Code) ▪ Roo (VS Code) ▪ Augment Code (VS Code, JetBrains, Cursor) Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
13. CLIs ▪ Codex CLI (OpenAI) ▪ Claude Code (Anthropic) ▪ GitHub Copilot CLI (GitHub) ▪ Gemini CLI (Google) ▪ Cline CLI Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding ▪ ForgeCode ▪ Aider ▪ Plandex ▪ GPT Engineer ▪ Smol Developer
14. Behind the scenes Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
15. Expectation management… ▪ “AI is useless. It’s like an extremely dumb intern, it always makes the same mistakes, and I have to tell it everything over an over again.” ▪ Of course! ▪ LLM’s don’t have memory or persistence. Only their context is relevant. Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
16. Expectation management… ▪ LLMs are a statistical melting pot of training data ▪ LLMs are trained on a lot of freely available source code ▪ What code is freely available (and thus trained on), and what code is NOT freely available (and not trained on?) ▪ Libraries, samples, (unfinished) pet & side-projects vs. real enterprise applications, products and your codebase Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
17. Expectation management… ▪ AI is just a tool ▪ Tools around AI are different tools ▪ We need to understand how it works to use it correctly ▪ We need to use the correct tool for the job Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
18. Pair Programming? ▪ Driver ▪ writes code ▪ Observer / Navigator ▪ reviews on the go ▪ considers “strategic” decisions as they happen ▪ Similar to aviation: Pilot Flying & Pilot Monitoring Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
19. Pair Programming? ▪ Roles are switched regularly ▪ Different tools for different roles ▪ CLI -> AI is the Driver ▪ IDE Integration / Chat: AI should be the Navigator ▪ Beware of the vibe….! Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
20. Potential problems Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
21. Potential problems ▪ AI as driver ▪ Overwhelming fast pace ▪ Human is too slow to keep up generated code ▪ Do we really want to only review new code? Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
22. Potential problems ▪ Tool & model selection ▪ Switching tools / models with same instructions yields very flaky results ▪ Model doesn’t know our guidelines ▪ Models sometimes change random stuff Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
23. Possible solutions Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
24. Possible solutions ▪ Prevent reviewing becoming overwhelming ▪ Organizational solutions where possible ▪ e.g. automatic code formatting ▪ Limit arbitrary code changes to the minimum ▪ Let the model “explain” its rationale with the code Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
25. Possible solutions ▪ Provide exact and on-point instructions (aka Prompt engineering) ▪ Make use of templates ▪ Adjust your instructions all the time ▪ ➔ Context Management Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
26. Context is king Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
27. Context is King ▪ LLMs love Markdown ▪ README.md, AGENTS.md, CLAUDE.md, Contributing.md etc. ▪ Memory ▪ Tasks, Todos, Guidelines etc. need to be memorized ▪ Dependencies ▪ Model needs to know the APIs or else… ▪ it can / will hallucinate calls https://github.com/josix/awesome-claude-md Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
28. Context is King ▪ Provide a Manifesto / Constitution for a project ▪ Defines rules that must be followed always ▪ Idea: Spec-driven-development? ▪ maybe try GitHub Spec Kit https://github.com/github/spec-kit Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
29. Context is King ▪ MCP (Model Context Protocol) provides access to information (context) as tool calls ▪ but... Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding https://julsimon.medium.com/why-mcps-disregard-for-40-years-of-rpc-best-practices-will-burn-enterprises-8ef85ce5bc9b
30. Context ist King: MCP ▪ Some Examples: ▪ Context7 ▪ Playwright ▪ GitHub (issues, source) ▪ Atlassian MCP Server (Jira, Confluence) ▪ Apidog ▪ Sequential Thinking MCP Server (well… ) ▪ … Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
31. Context is King ▪ Claude now supports “Skills” ▪ Skill: ▪ Selective context / instructions as Markdown file in a separate folder, supported by scripts (tools) where needed ▪ e.g. Awesome Thinking https://github.com/PawelGerr/ai-plugins https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
32. Context is King ▪ Sub-Agents help with context management ▪ Sub-Task does not clutter context window of main agent https://anthropic.mintlify.app/en/docs/claude-code/sub-agents https://code.visualstudio.com/docs/copilot/customization/custom-agents Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
33. Context is King ▪ Sub-Agents can be instructed much more specifically ▪ Backend- and / Frontend-Specialists ▪ WPF to Angular rewriter ▪ WCF to Web API converter ▪ Code-Reviewer ▪ Performance reviewer / Security reviewer ▪ Test-Engineer… Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding https://anthropic.mintlify.app/en/docs/claude-code/sub-agents https://code.visualstudio.com/docs/copilot/customization/custom-agents
34. Conclusion Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
35. Conclusion ▪ Try to stick to one tool ▪ but maybe don’t completely commit to it - yet ▪ For new models: try them out ▪ adjust prompts ▪ compare results ▪ upgrade when results are reliable Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
36. Additional things to consider ▪ Feature sets are still very different ▪ Model availability is important ▪ Share experiences with your team ▪ We’re all still in the learning phase! Wie hole ich das Beste aus AI-Assistenten heraus? p"AI"r Programming statt Vibe-Coding
37. Thank you! Sebastian Gingter sebastian.gingter@thinktecture.com https://thinktecture.com/sebastian-gingter .NET Developers Conference 2025