Show notes
Today’s episodeToday I’m showing you how to build a company operating system, with Mikhail Shcheglov, CPO at OLX Classifieds.After 5 months of continuous building on OpenClaw and Hermes, his entire product team now runs on it.His knowledge graph covers 54% of the company's product, business and customer context. That is a number he tracks as a personal KPI. At that level the agent already makes backlog decisions. Stakeholders pitch feature requests to it before they’re allowed near a PM. It runs his email, his calendar, his recruiting funnel and his design system.He opens his IDE on camera and shows all of it. 2 findings cut against everything you’ve been told. Summarizing your meeting transcripts costs you 20-25% recall, so he stores every single one raw. And letting Hermes write its own skills off repeated tasks produced a 31% accuracy lift in controlled testing.We also get into something heavier than architecture.What happens to PM headcount when one PM covers four domains? And what does a CPO actually screen for now when hiring?Don't miss....Brought to you byBolt.new - Ship AI-powered products 10x fasterProduct Faculty - Get $150 off their #1 AI PM Certification: code AAKASH150Customer.io - Send smarter messages using your product dataLand PM Job - 12-week live course to master the PM job searchViktor - Use $100 in starting credits to get 5x more done with this AI employeeIf you want access to my AI tool stack including Airtable, Speechify, Descript, Magic Patterns, Linear, Dovetail, Arize and Mobbin, that's $27,000 of value for $150, grab https://bundle.aakashg.com/. Key TakeawaysContext coverage is a CPO level KPI - Mikhail tracks what percentage of the company's industry, business model and customer knowledge his agent actually holds. It sits at 54%. That is enough for it to operate like a junior to mid PM and make backlog calls. At 70 to 90% he expects strategy level work.The real problem AI solves is knowledge leakage - A domain expert leaves and takes five years of context with them. Every company has this hole and almost nobody measures it. One store of business, customer, product and technical knowledge closes it, and the better your AI knows that context the more you can hand it.Do not summarize your transcripts - Summarization cost them 20 to 25% recall. You lose the granular detail where the answer usually lives, and you force every conversation into a template it was never shaped like. Store everything raw.Memory needs three layers, not one - A knowledge graph for structure, a vector database for fuzzy retrieval, and raw daily transcripts in MD files. Exact keyword matching fails on most real queries because real questions are ambiguous. The vector layer carries the load.Auto generated skills lifted recall by 31% - Hermes watches what you keep asking for and decides on its own that a skill is worth writing. Tested across five core topics with ten questions each, control group against treatment group. Plus 31% accuracy.Imperatives matter more than prompts - Their rules file runs 700 lines. No fabrications. Think before you act. Facts over guesswork. And a ban on what he calls fake helpful, where the agent can't do the thing so it explains how you could do it yourself.CLAUDE.md stays short, SOUL.md goes long - CLAUDE.md holds under 100 lines and carries the highest priority. SOUL.md runs 800 and sits second. Some of those 800 lines contradict each other and it still produces his most accurate output, because every imperative gets tested against real queries.Make the agent the gatekeeper - Stakeholders are trained to pitch the agent first. It asks clarifying questions, checks the request against priorities already set, declines politely if it doesn't clear the bar, and routes it to the right PM if it does. The org chart is mapped internally so it knows who owns what.Half of PM time is process, not thinking - Weekly reports, stakeholder updates, demos. Delegate that layer and one PM does the work of two, pointed entirely at discovery. He now runs one PM across three or four customer facing domains, and only keeps dedicated owners on monetization and search.Own the agent yourself or lose the advantage - Feedback arrives daily and he pushes changes from his phone straight into the repo. Hand it to an AI ops hire or an engineering team and you keep the tool but lose the speed. Nobody without skin in the game iterates fast enough.Related contentI’ve already created all of the resources Mikhail mentioned in this episode. But I never mapped them all to one place. That changes today in this edition. PFA the map sorted by the parts you want to build.Memory first, I Built You Memory for Claude Code, Hermes and OpenClawOnce you've got a memory layer, put it in version control. GitHub for PMs has the 3-repo setup I use for skills and evalThe Hermes Agent Guide for PMs is a 20-minute setup and a 30-day rollout with persona and skill templates includedFor OpenClaw, start with Mahesh Yadav on becoming a Builder PMFor the org-wide version, watch my episode with Jiaona Zhang, CPO at Laurel, on building a Company OSMy conversation with Hannah Stulberg of DoorDash on building a Team OS is the same thing one floor down, with a free starter kit👨💻 Where to find Mikhail Shcheglov:LinkedIn: https://www.linkedin.com/in/scheglovm1/Substack: https://corpwaters.substack.com/👨💻 Where to find Aakash:Twitter: https://x.com/aakashguptaLinkedIn: https://www.linkedin.com/in/aagupta/Newsletter: https://www.news.aakashg.com/If you want to advertise, email [email protected]. This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.news.aakashg.com/subscribe