Press Space to continue
Raw transcripts preserve more useful detail than summaries
Press Space to continue
@mshcheglov1 assumed he had to summarize meeting transcripts before feeding them to his agent. He tested it. Summarizing made recall 20 to 25% worse. Everyone's instinct here is the same. Raw transcripts are noisy, full of tangents and half-finished sentences, so you clean them up first. That instinct is expensive. He gives two reasons it fails. The first is granularity, because the detail that answers a question six weeks later is exactly the kind of thing a summary drops as noise. The…
Aakash Gupta@aakashguptaAug 29
Agents are building most of the features at OLX Uzbekistan. Mikhail Shcheglov (@mshcheglov1) is the CPO there, and over the last 5 months he built an entire company operating system on top of OpenClaw and Hermes. He doesn't read his email anymore. The agent does, and pings him only when something's urgent. It also screens stakeholder feature requests before they reach a PM, and it updates the Figma design system on its own. Then he open sourced the repo. We discuss: 🔶 The knowledge graph his whole company runs on 🔶 Why summarizing transcripts cost him 25% recall 🔶 The agent that kills bad feature requests 🔶 The "board skill" that pokes holes in your deck 🔶 Why the CPO should own this, not an AI ops hire 🔶 The one question that spots AI-native PMs Watch the full episode 👇 https://youtu.be/zocznD8Z-6k
View