Y Combinator CEO Garry Tan has shared how he built a personal AI “second brain” around a 100,000-page structured knowledge base, more than 100 composable skills, and a lightweight framework that routes tasks to the right workflow. In his description, the system is not a chatbot or a search box. It behaves more like an operating system that keeps running, updating, and reusing context across work and reading.
A 30,000-word book mirror tied to real life
One of Tan’s main examples is a “book mirror” for Pema Chödrön’s When Things Fall Apart. The system extracted all 22 chapters and ran two parallel processes for each one: summarizing the author’s ideas and mapping those ideas onto Tan’s own life, work, reading history, and prior conversations. The output was a 30,000-word report with two columns for each chapter, one focused on the book and the other on how it connected to events and patterns in his own experience.
He said the process took about 40 minutes. Tan added that he has already used the same method on more than 20 books, including Designing Your Life, Siddhartha, and Steppenwolf. The point, as he framed it, is not just a better summary. Each new book enters the same memory system, so later outputs can draw on the earlier ones.
From manual workflows to reusable skills
Tan said the first version of the book mirror performed poorly and even contained factual errors about his family background. He responded by adding a mandatory verification layer and assigning different models to different checks: Opus 4.7 1M for precision errors, GPT-5.5 for missing context, and DeepSeek V4-Pro for passages that sounded too generic.
That process became part of what he calls “skillification,” turning repeated manual work into reusable skill files with triggers, edge cases, and tests. At the center of that loop is a meta-skill called Skillify. When Tan encounters a repeatable workflow, Skillify inspects what just happened, extracts the pattern, writes the skill definition, and registers it with the system’s parser. In his setup, skills do not just execute tasks. They can also generate new skills.
Meeting prep built from accumulated context
Tan also described how the system handled preparation for a YC fireside chat featuring Demis Hassabis. In less than two minutes, it assembled Hassabis’ accumulated profile from articles, podcast transcripts, and notes; pulled together his public AGI timeline view; summarized points from Sebastian Mallaby’s biography; listed research priorities; cross-referenced them with Tan’s own public AI views; and drafted demo scripts and conversation hooks for the discussion.
He said each person in the system has a dedicated page with timelines, status fields, unresolved threads, and a score. Every meeting produces a transcript and structured summary, and the system then updates the pages of the people and companies mentioned in that meeting. Tan described this as “entity propagation,” a mechanism that keeps the knowledge base growing after every interaction.
Fat skills, fat code, thin framework
Tan summarized the architecture with the phrase “fat skills, fat code, thin framework.” OpenClaw acts as the execution layer and router, deciding which skill should handle a request without carrying domain knowledge about books, meetings, or founders. The real logic sits in the skills layer, where he now says there are more than 100 Markdown-based skill files covering tasks such as meeting-ingestion, enrich, media-ingest, and perplexity-research.
The data layer is the 100,000-page knowledge base itself, which stores linked records for people, companies, books, articles, podcasts, videos, meetings, and ideas. Models are treated as replaceable components inside the system: Opus 4.7 1M for accuracy, GPT-5.5 for recall and full extraction, DeepSeek V4-Pro for creative tasks, and Groq plus Llama for speed. Tan’s argument is that the model is only the engine, while the rest of the stack determines what the vehicle can actually do.
The stack is open-source
Tan said he has open-sourced the system, including the skills, the book mirror workflow, the cross-modal evaluation framework, the Skillify loop, the parser architecture, and more than 30 installable skill packs. For people building their own version, he suggested starting with a thin framework, creating a personal knowledge base, trying one concrete task first, then extracting the working pattern into a reusable skill.
According to Tan, the system now runs with 100 scheduled tasks operating continuously, handling meeting ingestion, email sorting, and updates to the knowledge graph.