Culture & Ideas6/10
Garry Tan Describes Using AI Agents to Mirror Pema Chodron's Book
Y Combinator CEO Garry Tan publishes an article on meta-meta-prompting and personal AI operating systems, describing how he had an AI agent summarize Pema Chodron's book and map each chapter to his own life. The piece is part of a series on agent architecture and open-source tooling.
Original post · 12 min read
X ArticleMeta-Meta-Prompting: The Secret to Making AI Agents Work
People keep asking me why I am spending my nights coding til 2AM. I have a job and a big one, as CEO of Y Combinator. We help thousands of builders a year to create their dreams of building real startups with real revenue that grow fast.
In the last 5 months, AI made me a builder again. Late last year, the tools got good enough that I went back to building. Not toy projects. Real systems that compound. I want to show you, with specific examples, what personal AI actually looks like when you stop treating it as a chat window and start treating it as an operating system. And I give it away as open source and in articles like this because I want you to speed up with me.
This is part of a series: Fat Skills, Fat Code, Thin Harness introduced the core architecture. Resolvers covered the routing table for intelligence. The LOC Controversy was about how every technical person just multiplied themselves by 100x to 1000x. Naked models are stupider argued that the model is the engine, not the car. And the skillify manifesto explained why LangChain raised $160M and gave you a squat rack and dumbell set without a workout plan, and then gave you that workout plan you needed.
The Book That Read Me Back
Last month I was reading Pema Chödrön's When Things Fall Apart. It's 162 pages, 22 chapters on Buddhist approaches to suffering, groundlessness, and letting go. A friend recommended it during a hard period.
I asked my AI to do a book mirror.
What that means concretely: The system extracted all 22 chapters of the book, and then, for each chapter, ran a sub-agent that did two things simultaneously: summarized the author's ideas, and then mapped every idea to my actual life. Not generic "this applies to leaders" pablum. Specific mapping. It knows my family history (immigrant parents, dad from Hong Kong and Singapore, mom from Burma). It knows my professional context (running YC, building open-source tools, mentoring thousands of founders). It knows what I've been reading, what I've been thinking about at 2am, what my therapists and I are working on.
The output was a 30,000-word brain page. Each chapter rendered as two columns: what Pema says, and how it maps to what I'm actually living through. The chapter on groundlessness connected to a specific founder conversation I'd had the week before. The chapter on fear mapped to patterns my therapist had identified. The chapter on letting go referenced a late-night session where I'd written about the creative freedom I'd found this year.
The whole thing took about 40 minutes. A $300/hour therapist reading this book and applying it to my life couldn't do this in 40 hours, because they don't have the full graph of my professional context, my reading history, my meeting notes, and my founder relationships all loaded and cross-referenceable.
I've done this with over 20 books now: Amplified (Dion Lim), Autobiography of Bertrand Russell, Designing Your Life, Drama of the Gifted Child, Finite and Infinite Games, Gift from the Sea (Lindbergh), Siddhartha (Hesse), Steppenwolf (Hesse), The Art of Doing Science and Engineering (Hamming), The Dream Machine, The Book on the Taboo Against Knowing Who You Are (Alan Watts), What Do You Care What Other People Think (Feynman), When Things Fall Apart (Pema Chodron), A Brief History of Everything (Ken Wilber), and more. Each one gets richer because the brain gets richer. The second mirror knew about the first. The twentieth knew about all nineteen.
How Book-Mirror Got Better Through Iteration
The first book mirror I did was terrible. Version 1 had three factual errors about my family. It said my parents were divorced when they weren't. Said I grew up in Hong Kong when I was born in Canada. Basic stuff that could have damaged trust if I'd shared it.
So I added a mandatory fact-check step. Every mirror now runs cross-modal evaluation against known facts in the brain before it ships. Opus 4.7 1M catches precision errors. GPT-5.5 catches missing context. DeepSeek V4-Pro catches when something reads as generic.
Then I upgraded to deep retrieval with GBrain tool use. The original version was good at synthesis but weak on specificity. Version 3 does per-section brain searches. Every right-column entry cites actual brain pages. When the book talks about dealing with difficult conversations, it doesn't just synthesize general principles. It pulls from my actual meeting notes with specific founders who were having tough conversations with co-founders. Or that idea I had on a Thursday hanging out with my brother James. Or the IM chat I had with my college roommate when I was 19. It's uncanny.
This is what skillification (using /skillify in GBrain) means in practice. I took the first manual attempt, extracted the repeatable pattern, wrote a tested skill file with triggers and edge cases, and every fix compounded across all future book mirrors.
Skills That Build Skills
Here's where it gets recursive, and where I think the biggest insight is.
The system that run… continue on X ↗
In the last 5 months, AI made me a builder again. Late last year, the tools got good enough that I went back to building. Not toy projects. Real systems that compound. I want to show you, with specific examples, what personal AI actually looks like when you stop treating it as a chat window and start treating it as an operating system. And I give it away as open source and in articles like this because I want you to speed up with me.
This is part of a series: Fat Skills, Fat Code, Thin Harness introduced the core architecture. Resolvers covered the routing table for intelligence. The LOC Controversy was about how every technical person just multiplied themselves by 100x to 1000x. Naked models are stupider argued that the model is the engine, not the car. And the skillify manifesto explained why LangChain raised $160M and gave you a squat rack and dumbell set without a workout plan, and then gave you that workout plan you needed.
The Book That Read Me Back
Last month I was reading Pema Chödrön's When Things Fall Apart. It's 162 pages, 22 chapters on Buddhist approaches to suffering, groundlessness, and letting go. A friend recommended it during a hard period.
I asked my AI to do a book mirror.
What that means concretely: The system extracted all 22 chapters of the book, and then, for each chapter, ran a sub-agent that did two things simultaneously: summarized the author's ideas, and then mapped every idea to my actual life. Not generic "this applies to leaders" pablum. Specific mapping. It knows my family history (immigrant parents, dad from Hong Kong and Singapore, mom from Burma). It knows my professional context (running YC, building open-source tools, mentoring thousands of founders). It knows what I've been reading, what I've been thinking about at 2am, what my therapists and I are working on.
The output was a 30,000-word brain page. Each chapter rendered as two columns: what Pema says, and how it maps to what I'm actually living through. The chapter on groundlessness connected to a specific founder conversation I'd had the week before. The chapter on fear mapped to patterns my therapist had identified. The chapter on letting go referenced a late-night session where I'd written about the creative freedom I'd found this year.
The whole thing took about 40 minutes. A $300/hour therapist reading this book and applying it to my life couldn't do this in 40 hours, because they don't have the full graph of my professional context, my reading history, my meeting notes, and my founder relationships all loaded and cross-referenceable.
I've done this with over 20 books now: Amplified (Dion Lim), Autobiography of Bertrand Russell, Designing Your Life, Drama of the Gifted Child, Finite and Infinite Games, Gift from the Sea (Lindbergh), Siddhartha (Hesse), Steppenwolf (Hesse), The Art of Doing Science and Engineering (Hamming), The Dream Machine, The Book on the Taboo Against Knowing Who You Are (Alan Watts), What Do You Care What Other People Think (Feynman), When Things Fall Apart (Pema Chodron), A Brief History of Everything (Ken Wilber), and more. Each one gets richer because the brain gets richer. The second mirror knew about the first. The twentieth knew about all nineteen.
How Book-Mirror Got Better Through Iteration
The first book mirror I did was terrible. Version 1 had three factual errors about my family. It said my parents were divorced when they weren't. Said I grew up in Hong Kong when I was born in Canada. Basic stuff that could have damaged trust if I'd shared it.
So I added a mandatory fact-check step. Every mirror now runs cross-modal evaluation against known facts in the brain before it ships. Opus 4.7 1M catches precision errors. GPT-5.5 catches missing context. DeepSeek V4-Pro catches when something reads as generic.
Then I upgraded to deep retrieval with GBrain tool use. The original version was good at synthesis but weak on specificity. Version 3 does per-section brain searches. Every right-column entry cites actual brain pages. When the book talks about dealing with difficult conversations, it doesn't just synthesize general principles. It pulls from my actual meeting notes with specific founders who were having tough conversations with co-founders. Or that idea I had on a Thursday hanging out with my brother James. Or the IM chat I had with my college roommate when I was 19. It's uncanny.
This is what skillification (using /skillify in GBrain) means in practice. I took the first manual attempt, extracted the repeatable pattern, wrote a tested skill file with triggers and edge cases, and every fix compounded across all future book mirrors.
Skills That Build Skills
Here's where it gets recursive, and where I think the biggest insight is.
The system that run… continue on X ↗
♥ 3.8K · ⟲ 502 · 👁 1.6MView on X ↗
