Say What You Mean: Spec-Driven Development and the End of AI Guesswork
Video Description
Speakers: William Hurley and Matthew Clewley
You've been here: you describe what you want, the agent hands back a clean block of code that compiles, and then you realize it solved the wrong problem. The fix isn't a cleverer prompt. Coding agents are literal-minded pair programmers that need unambiguous instructions, not search engines.
This session is about spec-driven development: making a specification, not your chat history and not the code, the thing your AI agent builds against. We'll work through OpenSpec, a lightweight take on the idea. It keeps your current specs as the source of truth and gives each proposed change its own folder with a proposal, specs, design, and tasks. A change starts with one command, the agent drafts the artifacts, and you review them before any code is written, then apply and archive when it's done. You'll see how every change becomes deterministic and reviewable instead of a wall of code to rubber-stamp, plus the honest tradeoffs: more upfront thinking, more tokens, and the discipline to keep the spec alive instead of letting it rot.
You'll leave knowing how to write a spec an agent can build from, how to keep your specs and code in sync as the system changes, and how to fold this into the way your team already works. Bring a project you're stuck reverse-engineering from your own AI's output, and you'll see how to flip it around.
You've been here: you describe what you want, the agent hands back a clean block of code that compiles, and then you realize it solved the wrong problem. The fix isn't a cleverer prompt. Coding agents are literal-minded pair programmers that need unambiguous instructions, not search engines.
This session is about spec-driven development: making a specification, not your chat history and not the code, the thing your AI agent builds against. We'll work through OpenSpec, a lightweight take on the idea. It keeps your current specs as the source of truth and gives each proposed change its own folder with a proposal, specs, design, and tasks. A change starts with one command, the agent drafts the artifacts, and you review them before any code is written, then apply and archive when it's done. You'll see how every change becomes deterministic and reviewable instead of a wall of code to rubber-stamp, plus the honest tradeoffs: more upfront thinking, more tokens, and the discipline to keep the spec alive instead of letting it rot.
You'll leave knowing how to write a spec an agent can build from, how to keep your specs and code in sync as the system changes, and how to fold this into the way your team already works. Bring a project you're stuck reverse-engineering from your own AI's output, and you'll see how to flip it around.