Y Combinator CEO Garry Tan took to X to unveil his personal AI operating system, sparking widespread discussion among developers. He argues that most people still treat AI as a chat window, while the real leverage lies in building a continuously compounding 'second brain' around personal knowledge, workflows, and judgment. Tan not only shared the system architecture but also open-sourced all the code, inviting anyone to replicate his approach.
The 40-Minute Book Mirror: How a 30K-Word Personalized Note Is Made
Tan demonstrated the system using Pema Chödrön's 'When Things Fall Apart.' He asked the AI to perform a 'book mirror': the system extracted all 22 chapters, ran a sub-agent for each chapter, simultaneously summarizing the author's ideas and mapping every point to Tan's real life—including his immigrant family background, YC management context, recent readings, late-night thoughts, and even issues his therapist was addressing. The output was a 30K-word 'brain page' with two columns per chapter: one column for the original content, the other for personalized mappings. The entire process took about 40 minutes. Tan emphasized that a therapist billing $300/hour couldn't achieve this in 40 hours, because the system has cross-referenced his professional context, reading history, meeting notes, and founder network.
Tan has processed over 20 books this way, including 'Amplified,' 'The Autobiography of Bertrand Russell,' 'Designing Your Life,' 'Finite and Infinite Games,' 'Siddhartha,' and 'Steppenwolf.' Each book enriches the 'brain,' with the second mirror knowing the first, and the twentieth knowing all previous nineteen.
From File Cabinet to Nervous System: Automating a 100K-Page Knowledge Base
Tan maintains a structured knowledge base of roughly 100,000 pages. Each person has a page with a timeline, status bar, and score. Every meeting has a transcript and structured summary. Every book gets a chapter-by-chapter mirror. Every article, podcast, and video is ingested, tagged, and cross-referenced. The system uses 'entity propagation': after each meeting, it iterates through every person and company mentioned and updates their brain page with the discussion content. Tan calls this the shift from a 'file cabinet' to a 'nervous system'—the former just stores, while the latter connects, marks changes, and surfaces the most relevant information. When Demis Hassabis visited YC for a fireside chat, the system pulled in under two minutes Demis's full brain page, his public AGI timeline views, key parts of Mallaby's biography, his research priorities, cross-references with Tan's AI views, and a set of conversation entry points.
The Meta-Skill 'Skillify': Turning One-Off Tasks into Compounding Assets
Tan's core insight is 'Skillify'—a meta-skill for creating new skills. Whenever he encounters a repetitive workflow, he says 'skillify this,' and the system extracts the pattern, writes a tested skill file with triggers and edge cases, and registers it with the resolver. The system now has over 100 skills, each focused on one task but composable into complex workflows. For example, the book-mirror skill calls brain-ops for storage, enrich for context, cross-modal-eval for quality check, and pdf-generation for output. Improving one skill automatically improves every workflow that uses it.
Open-Source Architecture: Thin Harness + Thick Skills
Tan open-sourced the entire stack with a 'thin harness, thick skills' design. The harness (e.g., OpenClaw and Hermes Agent) is only a few thousand lines of routing logic, agnostic to specific tasks. The skills are thick, self-contained markdown files—now over 100—each with detailed instructions for a specific task. Models are interchangeable: Opus 4.7 1M for precision, GPT-5.5 for recall, DeepSeek V4-Pro for creativity, Groq with Llama for speed. Skills decide which model to call for each job.
Tan runs over 100 cron jobs daily, monitoring social media, Slack, and email. He thinks not about productivity but about compounding—every meeting, every book, every skill adds capability to the system that improves every hour. He believes the future belongs to individuals who build compounding AI systems, not those who only use centralized corporate AI tools. All code, skills, the book-mirror pipeline, cross-modal eval framework, skillify loop, resolver architecture, and over 30 installable skillpacks are freely available on GitHub.