Spec-Driven Development: Structured AI Coding Beyond Vibe Coding

Key Quotes

"You know, this AI space is moving so fast, there are so many new concepts and so, and you know, and some reusing of old concepts too."

Marc Brooker explains that the rapid advancement of AI introduces many novel ideas, but also reintroduces or recontextualizes existing concepts. This highlights the dynamic nature of the field, where innovation often builds upon or repurposes prior knowledge.

"Spec-driven development is this idea that, hey, instead of starting with writing a program with the AI, you know, the AI model and and the developer are going to work together to write a specification, just a a description of what the program does and step through that, and then write some requirements of like, well, hey, I want to build this in Python, I want to use these libraries, I don't want to use these libraries, I want it to work on these operating systems and so on."

Marc Brooker defines spec-driven development as a collaborative process between a developer and an AI model. This approach prioritizes creating a detailed specification and requirements document before writing code, contrasting with prompt-by-prompt AI coding.

"I think we all know that you don't always know that completely upfront; you discover that as you go. And so, you know, I don't want to imply that these specs are an immutable artifact. What you're going to do as you develop and as you test and as you talk more to your customers is keep going back and updating the spec and saying, well, here's what I've learned, here's some new user stories, here's some new requirements."

Marc Brooker emphasizes that specifications are not static documents. He explains that as development progresses and feedback is gathered, specifications should be iteratively updated to reflect new learnings and evolving requirements.

"Imperative programs are like, do this, and then do this, and then do this, do this loop. Add these numbers together and so on. And that's what a lot of, you know, is especially the Python code, the C code, the Rust code that people are writing. It's like, and then there's this other approach to coding which is more declarative, which is, I want this result. You know, it's more like, go off and figure out how to do this result."

Marc Brooker differentiates between imperative and declarative programming. He describes imperative programming as a step-by-step instruction set, common in languages like Python, while declarative programming focuses on specifying the desired outcome, leaving the execution details to the system, as seen in SQL.

"I think that tasks file is somewhat an artifact of the limitations of today's LLMs and their ability to do things over long time horizons. And so forcing the LLM to say, make a plan, make a plan to achieve this, really helps it kind of do things in bite-sized chunks and not get lost along the way."

Marc Brooker explains that the "tasks file" in Kiro is a way to manage LLM limitations. He notes that by breaking down development into smaller, planned tasks, LLMs are better able to execute complex processes without losing track of the overall goal.

"The fundamental skill of, you know, listening to your users or or thinking about your business and crisply writing down what you want to achieve, that's been a great skill for programmers to have for 40 years. It's going to be a great skill for, you know, for folks to have in another 40 years. I don't see that changing. And so that's the core of spec-driven development is getting good at at crisply explaining what you know what a program should do."

Marc Brooker identifies the core of spec-driven development as the enduring skill of clearly articulating desired outcomes. He asserts that this ability to listen to users and precisely define program objectives is a foundational programming skill that will remain valuable for decades to come.