A conference talk by Anton Arhipov (JetBrains) on spec-driven development with AI agents. The core idea is formalizing requirements into structured markdown documents (proposal → plan/design → task list) before handing off to an AI agent, rather than using ad-hoc prompts. Key insights include: context is the most precious resource when working with agents; persisting specs in files survives agent crashes; agent 'skills' (guidelines files) dramatically improve code quality by encoding architectural conventions; task lists serve as progress trackers and review artifacts. The talk demos OpenSpec, a community toolkit that automates this workflow, and compares it to similar tools like Kiro IDE and SpecKit. Real-world lessons from generating a Spring Boot e-commerce app show that without architectural skills/guidelines, agents produce poor layered architecture and duplicated exception handling. The speaker argues spec-driven development bridges the gap between vague prompt-driven coding and predictable, reviewable software generation, though tooling is still maturing.
Sort: