HandyProductEngine

I run an AI practice with several client apps, and the bottleneck was never the code. It was the product management: signal scattered across boards and inboxes, no consistent way to prioritize, and half the shipping decisions living only in my head.

I wanted a system that ran my own product process end to end, so I built one and pointed it at itself.

It runs as two loops. A build loop turns customer signal into a triaged roadmap, has AI agents draft the feature spec and build it, and puts a human on the gate at both the spec and the pull request. A knowledge loop distills every build into shared standards and reusable playbooks that ground the next spec and get applied across every app the practice manages. GitHub is the source of truth, and every agent write is a pull request.

I put the human on the two highest-leverage gates, approving the spec and reviewing the PR, and automated everything in between. Judgment stays where it matters; the rest runs on its own.

Every change is a pull request into GitHub, so the repo is literally the product's memory: auditable and revertible. And I kept standards and playbooks at the fleet level rather than per app, because the real payoff is cross-project leverage, where one learning improves every similar product I run.

The value was never a single agent. It was the compounding loop, and the standards and playbooks, not the code generation, turned out to be the durable advantage.

The other lesson was about constraints. Forcing every change through a reviewed pull request is what kept the quality bar honest.