A DoorDash PM built a Claude Code system that supports 20+ people.
Hannah Stulberg. Ex-Google APM. 1,500+ hours in Claude Code. She calls it a “Team OS.”
Most PMs use AI as a personal productivity tool. One person benefits. Everyone else on the team still waits for context, still searches for the right query, still asks the PM to pull the number.
Hannah did something different. She built a shared repo that every function on her team can access. Data scientists pull their own metrics. Strategy leads grab their own customer call summaries. Engineers check analytics without filing a ticket.
The setup takes about 30 days. Here’s the architecture:
1. A root CLAUDE md file (under 500 tokens) with the doc index, team roster, Slack IDs, and GitHub handles.
2. Nested CLAUDE md files in every major folder. Navigation indexes only. 3% context per query.
3. Folder ownership assigned by function. PM owns product. Data scientist owns analytics. Strategy owns customer calls.
4. An analytics layer with metric definitions, SQL queries, and dashboard links, loaded progressively by depth.
5. Shared skills: customer call summaries, PR creation, weekly synthesis. Everyone runs the same tools.
6. Verified playbooks for recurring analyses (funnel drop-off, churn investigation). Claude follows the same steps your analyst would.
7. A launch gate: no feature ships until its metrics, queries, schemas, dashboards, and playbooks are checked into the repo.
The result is that Hannah went from being the team’s bottleneck to being the person who built the system that made the bottleneck disappear.
The barrier to starting is psychological. The terminal feels scary for about an hour. Then it feels normal.
Resources to build yours:
1. Full podcast episode where she walks through every step:
https://lnkd.in/gVi3PQ-U
2. Newsletter deep dive w starter repo:
https://lnkd.in/gi3vNyzR
3. My PM OS:
https://lnkd.in/gsH9RSfJ
Build the system that makes you unnecessary to the system.