.NET Rocks! GitHub Spec Kit with Den Delimarsky
6 snips
Oct 30, 2025 Dan Delamarsky, a Principal product engineer at Microsoft, shares insights on the GitHub Spec Kit and specification-driven development. He discusses how breaking down tasks into granular specs helps LLMs generate quality code more effectively. Delamarsky emphasizes the importance of refining specifications to keep AI models focused and manageable. The conversation touches on using test-driven loops to complete generated code, the evolving role of developers in overseeing AI outputs, and the challenges of adopting this new approach in software development.
AI Snips
Chapters
Transcript
Episode notes
Specs Separate Requirements From Tech Stack
- Spec-driven development detaches requirements from implementation so LLMs can generate multiple tech-stack variants from the same spec.
- This lets teams compare options quickly and choose the best implementation without rewriting requirements.
Use LLMs To Find Spec Blind Spots
- Use the LM to 'Clarify' your spec by asking it to probe blind spots and missing assumptions.
- Iterate this questioning until the spec contains the detail the model needs to avoid guessing.
Decompose Specs Into Consumable Tasks
- Break large specs into small tasks so models stay within their context window and complete work reliably.
- Task-level decomposition enables staged, test-driven generation and parallel sub-agents.