How to Use Spec-Driven Development for Production Workflows

Specs are the primary contract for generation of code.

Why?

AI as intern — one small deviation can result in very different results. Get planning and product involved.

Can't the latest frontier models do everything??

Let's keep a few things in mind:

• Too much context!! Models get confused.
• AGENTS.md / steering — keep it targeted.
• Use skills — create specs and implement, and agents can help with the implementation plan.
• Too much trust!! Are we code reviewing? We are the human in the loop, it's important.
• Set up AI code reviews as well / task rabbit etc.

Don't get off the intended target — outcome divergence.

Speed over maintainability — what patterns are being created?

History lesson — Alexa at AWS, Kiro, first as VS Code fork.

Iterating on spec-driven development

How do you do this stuff WITHOUT Kiro?

Tell the AI IDE what it should do:

• include the following
  • user requirements
  • design document from that
  • take both of those and create implementation details
  • Spec-kit / Open Spec / BMad

Can use Kiro spec mode in a brownfield application.

EARS — Easy Approach to Requirements Syntax.

requirements -> design phase

Review the markdown, approve, move to implementation. Implementation phase can be out of order.

Create an MVP from the steps — take a vertical slice, pedantically.

• Keep it short and tight, continue to review.

MCP

Isn't MCP dead? It's not as popular now… maybe still valuable for spec-driven development?

Switching to CLIs is pretty common.

MCP <-> Spec Driven Development

• specs can be pulled from project management service
• pulled from PM — take app/PoC and reverse-engineer it
• will follow rules set forth in your steering files

Property-based tests — testing to run regression tests against the requirements.

"Take all these steps and create the four that I can use to prove that this works."

Conceptually, does it work as I expect?