Building an AI toolkit for a front-end team

Why build our own

General-purpose AI coding tools don't know our standards: Sable, our ADRs, our accessibility expectations. I wanted tools that review and build code the way our team would, so they catch the things we actually care about.

What's in it today

• Spec-driven development workflow (5 agents). Takes a feature from requirements to design to tasks to implementation, with a specialized agent for each stage.
• Pull request review workflow (8 agents). Each agent looks at one concern, including a dedicated accessibility reviewer, so reviews are specific instead of generic.
• QA skill. Runs build, lint, and tests and reports problems before a human reviewer looks at the change.

The workflows started on my team and have been shared with the wider engineering org.

Guardrails

Every change still has a human owner. The agents make review faster and more consistent. They don't approve anything on their own.

What's next

• Checking that risky changes are behind feature flags
• Finding and drafting missing unit tests
• Suggesting Datadog RUM checks to watch after a deploy
• Estimating a pull request's blast radius
• Lighthouse-style performance checks that flag regressions before production