Adopting AI-assisted development without losing the plot

Situation

The team wanted the speed AI-assisted development promises, but without the failure mode that comes with it: piles of generated code no one fully understands and no one can safely change.

What I did

I rolled out a spec-first workflow team-wide: specifications written in OpenSpec, then implemented by Claude Code, with alternate harnesses used for specific tasks. I used multi-agent orchestration to parallelize implementation and review, and applied DSPy to systematically optimize and evaluate the prompts behind the platform’s AI-powered features. I also built a full retrieval-augmented-generation pipeline powering customer-facing chat, with model routing by use case to keep both quality and token spend in check.

Result

The spec-first workflow became a primary driver of the release-cadence gains from the platform turnaround, letting the team scale throughput without adding headcount, while keeping the system legible enough to change with confidence.