Spec-Driven Development: AI Assisted Coding Explained

This is exactly the traditional implementation coding - the waterfall model. So, we are back to the original development model here, but with way stupider workers - AI agents. Nice job.

Another great video! Thanks!

Please tell me, do you have to write mirrorred text on that board? How does it work

You forgot about plan-first vibe-coding We do rounds of planning back and forth until we reach the desired plan and only then we ask for the implementation

awesome very well explained

He needs to talk about vibe spec development as well. Telling the agent to create 10 specs for your project and then implementing it without any oversight

TLDR: Use specs instead of blind vibe coding. Very valuable, much wow.

writing good requirements is actually much harder than writing code. a good engineer knows how to build a system from vague requirements by interacting with business stakeholders.

Is it possible to use Gherkin approach withe Spec-driven development. Because that would be cool and probably the easiest way to do software. Why do I need to worry myself about codes like 401 and 200 and this is quite low level of abstraction what the software need to do.

I always start dividing the app into Epics and stories. The first epic is always about foundations. The LLM will create detailed plan for each story and reviewed by human. Coding only after the plan is ready.

I don’t get it. When you prompt an LLM (e.g., “add user login authentication”), it can already generate clear instructions or even the full blueprint—file structure, frontend, backend, and database. There’s no need to overcomplicate things.

I think documentation is better with LLM, but completely building something based on that documented spec might cause issues when there’s changes based on user feedback. Because AI will completely change what has already been built! If it’s possible to change that exact spot where we need to update or the feature that needs modifications, then I think this would be great!

As a software engineer I was playing with Claude code, then I noticed that it does better when we separate planning and execution, then it does even better when we first design, then plan, then implement and so on, turns out I independently invented spec-driven development :)

Small suggestion: On both <happy> decision boxes, you should label the output paths with "Y" or "N" so it is clear of the decision result. Also, you may want to always have the Y path on top (N on bottom) for consistency. Possibly overlay with text Y/N on top of video?

Yes, this is how the theory of software development has been since the 50's. All I want is someone to show me how it works in practice.

Back when humans were writing code, it seemed impossible to get specs/requirements. "Just implement something" I was told with some vague hand-waving, "if we don't like what you make you can just do it again". Now that machines are writing the code I'm being told "you need to give the poor LLM a proper spec so it can do its work". Damn machines get more respect than I ever did as a dev.

This offers excellent insight: when you have a team of highly intelligent employees capable of infinite iteration, DevOps Scrum proves to be a perfectly suited practice. By combining Scrum’s lightweight rituals with a heavy emphasis on documentation, you ensure that AI—that "tireless, zero-cost, and flawlessly executing" employee—always possesses full contextual awareness.

I think it’s more about test driven development. Specify the test and testing skills. The test takes care of itself.

For people commenting about the first line, interpret his words in this way, it will make sense --> "Before, writing and reviewing code ate up most of our precious brain-time, but now, most of our brain-time goes into knowing how to effectively convey and collaborate with AI to build what we want - and that's exactly why we need spec-driven development (SDD) - because traditional software development techniques were catered for humans, and just like how we are making everything AI-friendly, it's high time we evolve into SDD"

What if you ate forced to develop, operate and maintain software in a fully airgapped environment? It's very often that content may not been send to any cloud based services. The llm executable on local services are limited... Arnt they?