The problem
This came from a real bottleneck I hit as a PM. As developers increasingly ship with coding agents, engineering gets dramatically faster — and product work becomes the constraint. Everyone expects the PM to know what to build next, faster, but discovery, roadmapping, and writing PRDs still take days, with the signals scattered across Notion, Linear, Slack, and meeting notes.
AI should close that gap, but most “AI setups” are just prompt collections. Context is scattered across chats, output quality drifts, and nothing compounds — every session starts from zero, and none of it carries how you or your team actually work.
What I built
A skill-based operating system that turns AI agents from autocomplete into a reliable product-work partner. Built first in Cursor, my team's tool of choice, and portable across AI tools.
- Three layers: governance (what the AI runs autonomously vs. what needs human sign-off), context (how the company, team, and person actually work), and skills (each repeatable PM job as a versioned, self-improving skill).
- Skills: discovery, PRDs, roadmap passes — plain Markdown files in a repo the agent loads on demand.
- Compounding: meta-skills create, evaluate, log, and improve other skills, so every run enriches the system instead of starting from zero.
- Stack: Cursor (runtime) + Markdown repo + Notion (skill registry, output log, knowledge base) + MCP (Notion, Linear, Slack, GitHub).
- Portable: the same skills and context run across AI tools — Cursor at work, or Claude in private.
Why it matters
The system compounds. Meta-skills handle the upkeep — creating, evaluating, logging, and improving other skills — so every run enriches the substrate instead of starting from zero, and today's output becomes tomorrow's context.
The sharpest lesson: when agents can write the code, the leverage moves to context — a PM contributes to the product through clear, structured context, not syntax. It's a working system, and the clearest proof of context-engineering and harness literacy: tool loops, scaffolding, context management.
The lesson / where it's going
The intelligence lives in the system, not the individual. The leverage in AI isn't a better prompt — it's a structure that remembers how you work and compounds.