software design / ai engineering / ai agents
How to Design Software Systems With AI Before Writing Code
A working demo is only the beginning. Austin Starks explains the workflow he uses to design and build software systems with AI: write the design first, review it adversarially, and only then implement the code.
The design should examine existing patterns, cover the full stack, consider how the system scales, and include concrete code examples. A separate design review catches problems before implementation. Once the design is ready, a fast coding model can build it, followed by an adversarial code review.
The video closes with the checks that support shipping: unit tests, integration tests, end-to-end tests and CI/CD. Explore the linked course to learn more about building AI agents from scratch.
Transcript
0:00you're not using AI properly. And that's why your vibe coded app is breaking down at 3 a.m. Here's how you use AI to properly design software system. Number one, before you write a single line of code, you need to tell your AI to create a design.
0:14You explicitly need to tell it to look at existing patterns, create a system that scales, include code snippets, and to have the design span across the stack. And then, and this is the part where people most often skip, like you have a code review session, you need to have a design review session.
0:29Have Claude review its own design adversarially. Only then and after then can you use your favorite AI tool and just implement the code. And the implementation can be a cheap, fast model.
0:40You can use something like Composer 2.5. And then afterwards, you'll have your code review, you'll adversarially review all of your code, and you'll keep going until you get a ship ready verdict.
0:51By the way, this is exactly what I'm doing right now. And it's what I do all the time. By creating software systems this way, I can produce a high volume of complex software across the stack and ship it without manually reviewing every single line of code.
1:05Combined with things like unit tests, integration tests, end-to-end tests, and CICD, then you can build a high quality software system that literally repairs itself.
1:14You want to learn more about how to build production grade AI agents? Then click that follow button.
Join the conversation
Loading conversation…