I introduced a three‑layer AI operating‑system hierarchy – PM OS, Team OS, and Company OS – each with its own purpose, tools, and rating on Wade Foster’s AI fluency scale. The PM OS holds personal context and shortcuts (adoptive), the Team OS shares calls and decisions across the team, and the Company OS aggregates skills, transcripts, and email across the entire organization (transformative). I also laid out a separation rule (who needs the answer?) and a build‑order guidance that starts with Company OS when I’m constantly pinged for context, then Team OS for a small codebase, and finally PM OS for an individual Claude Code setup.
PM OS rating: Adoptive.
Company OS rating: Transformative.
Pilot result: the bot read twelve months of sponsor email, identified three sponsors, and after a single Slack line we produced a reviewed PRD and launched a live page.
On strategy I shared a seven‑step process (objective, users, super‑powers, vision, pillars, impact, roadmap) and emphasized that strategy work is intentionally messy – you’ll loop back constantly and treat the framework as a set of windows rather than the strategy itself.
Quote on messiness: “Everyone who has worked on strategy will tell you it’s messy chaos. And it’s supposed to be.”
Quote on “windows”: “The business model canvas. The lean canvas. The pitch deck your boss loves. He calls all of them windows into the work.”
Core strategy questions: “Why should we act, and why now? Who are we solving for, and who else solves it? How do we plan to win.”
Loop‑back reminder: “It never is. You loop back constantly.”
For a deeper dive into my AI‑product strategy this month see the n8n interview summary.
I published a monthly‑focused guide for product managers on how to adopt and operationalize new AI tools, built around a durable six‑step hierarchy that stays constant despite tool churn.
I define the five components of any real agent as model, knowledge, memory, tools, and guardrails – a rule I attribute to Mahesh Yadav.
I note that AI gross margins run 20‑60 % versus 70‑90 % for traditional SaaS.
I hand‑label about 100 real interaction traces to evaluate LLM judges, then compare raw agreement to true‑failure detection rate.
My content‑business workflow uses 51 distinct skills, organized by area‑folder and task‑skill.
I recommend a context window of one‑to‑two weeks of data, adding retrieval only for facts outside the model’s training.
I prioritize cost‑first optimization in the order: model routing → caching → pricing.
These decisions reflect my belief that a fixed hierarchy protects product teams from endless tool churn while keeping the focus on economic viability.