Software Engineering Radio - the podcast for professional software developers SE Radio 711: Scott Hanselman on AI-Assisted Development Tools
Mar 11, 2026
Scott Hanselman, Microsoft VP of Developer Community and longtime .NET educator, explains AI-assisted coding as a progression from autocomplete to agentic loops. He breaks down ambiguity in model behavior, the need for specific prompts and verification, when to rely on fundamentals versus agents, sandboxing and safety practices, and how to define testable success for model-driven work.
AI Snips
Chapters
Transcript
Episode notes
Minecraft One-Shot Demo Hid The Real Work
- A coworker one-shot a browser Minecraft clone, but Hanselman notes "Minecraft" contains a 50-page spec, so the demo relied on the heavy semantic weight of that name.
- He warns that removing that steering word yields chaotic results, showing one-shot demos can be misleading.
Learn Core Fundamentals To Steer Agents
- Learn fundamentals (HTTP, DNS, NAT, testing) so you can steer LLMs and judge their output instead of blindly trusting them.
- Hanselman compares it to driving stick: you don't need to always, but understanding the mechanics prevents blind dependence.
Approach LLMs With Strong Opinions And Clear Prompts
- Treat LLMs as either a junior engineer with infinite energy or a senior advisor depending on your expertise; come with strong opinions and explicit prompts.
- Hanselman recommends 'strong opinions, weakly held' and including prompt templates in coursework.