The team intentionally focused on using AI to make the group more productive, not just individuals. This prevents a scenario where everyone generates massive amounts of uncoordinated code and content, leading to chaos rather than progress.
Moving away from long-form documents, the team writes short 1-2 page PRDs focused on defining the customer problem. The primary output and center of debate is a functional prototype generated by an AI agent from that short document.
The team centralizes crucial context—product strategy, customer intelligence, meeting outcomes—into a repository of markdown files. This ensures all AI agents and team members pull from a single, up-to-date source of truth, making their outputs relevant and consistent.
A clear rule governs their shared skills repository: skills referencing specific code live within that code's repository. General-purpose skills go into a shared repo. This structure keeps skills co-located with their work, while personal experiments stay in private branches until proven valuable.
Instead of forcing everyone to maintain a complex local environment with all team contexts, they built "Orchestrator." This internal tool provides a unified interface to query any team's codebase and use their specific AI skills without deep setup, enabling casual cross-functional work.
To avoid shallow, AI-generated documents, their custom agent acts as a sparring partner. It conducts a turn-by-turn interview with the Product Manager, using a question bank to probe for tradeoffs and challenge assumptions before drafting the initial PRD.
Believing agents are a key user, the company built a tool that runs a suite of tests where an AI agent attempts to complete tasks using their product. This provides immediate, objective feedback on API design, SDK usability, and documentation clarity, replacing slow user testing.
The team reports significant velocity gains from AI but dismisses "3x" claims. While AI compresses research, coding, and testing, it doesn't solve the inherently human and time-consuming aspects of software development like debate, discovery, re-evaluation, and coordination.
Their workflow clearly separates human-centric tasks from automated ones. PMs are "human in the loop" for discovery, alignment, and defining API abstractions. Once those strategic decisions are made, the process of writing production-ready code is largely automated via AI agents.