Editor's Note: While most people still see AI as a smarter chat window, Y Combinator's current CEO, Garry Tan, has been attempting to transform it into a personal operating system.
The underlying structure of personal productivity in the AI era is changing: the model is just the engine, and what truly compounds is a whole system built around personal knowledge, workflows, context, and judgment.
In this system, every meeting, every book, every email, every relationship is no longer isolated information but is continuously written into a structured "second brain"; every recurring task no longer relies on ad-hoc prompts but is abstracted into reusable skills that are iterated on in subsequent work. In other words, AI is not just helping people complete tasks, but helping individuals productize, systematize, and infrastructuralize their way of working.
Of greater note, the author presents a personal trajectory different from centralized AI tools: future competitiveness may belong not only to those who can use AI but to those who can train a compounding AI system around their real life and work. While chatbots provide answers and search engines offer information, a true personal AI system will remember your background, understand your context, inherit your judgment, and grow stronger with each use.
This is also the most inspiring part of the article: the value of AI lies not in what it generates in a single instance, but in its ability to be a continuously accumulating, connecting, and improving neural system. For individuals, this may indeed be the true starting point of an "AI-native way of working."
Following is the original text:
People always ask why I spend my evenings coding, sometimes until 2 a.m. I have a job, and it's a very important job—I'm the CEO of Y Combinator. We help thousands of founders each year turn their dreams into real, revenue-generating, high-growth startups.
For the past five months, AI has transformed me back into a builder. By the end of last year, the tools were good enough for me to start building again. Not a toy project, but a system that can truly compound. I want to show you with concrete examples what it looks like when you no longer see personal AI as a chat window but as an operating system. I've open-sourced these things and written this article because I want you to speed up with me.
This is part of a series: "Fat Skills, Fat Code, Thin Harness" introduces the core architecture; "Resolvers" discusses the intelligent routing table; "The LOC Controversy" talks about how every tech person scales themselves 100x to 1000x; "Naked models are stupider" argues that models are just the engine, not the whole car; and the "Skillify Manifesto" explains why LangChain raised $1.6 billion, gave you a squat rack and dumbbells, but no training plan, whereas this article gives you that much-needed training plan.
The Book That Reads Me Backward
Last month, I was reading Pema Chödrön's "When Things Fall Apart." This 162-page, 22-chapter book delves into how Buddhism views pain, groundlessness, and letting go. A friend recommended it to me during a tough period in my life.
I had my AI perform a "book mirror."
Specifically, this meant: the system extracted the content of all 22 chapters of the book and then ran a sub-agent for each chapter, accomplishing two things simultaneously: summarizing the author's thoughts and mapping each viewpoint to my real-life situations.
Not the generic "this applies to leaders too" talk, but very specific mappings. It knew my family background: immigrant parents, father from Hong Kong and Singapore, mother from Myanmar. It knew my professional context: managing YC, building open-source tools, guiding thousands of founders. It knew what I've been reading lately, what I ponder at 2 a.m., what issues my therapist and I are working through.
The final output was a 30,000-word "brain page." Each chapter was presented in two columns: one for what Pema was saying and the other for how these contents mapped to what I was actually experiencing. The chapter on "groundlessness" connected to a specific conversation I had with a founder the week before; the chapter on "fear" mapped to some behavior patterns my therapist had pointed out; and the chapter on "letting go" referenced something I wrote late at night—about the creative freedom I found this year.
The whole process took about 40 minutes. Even a $300 per hour therapist couldn't have done this in 40 hours, even if they read the book and applied it to my life. Because they didn't have full loading and cross-referencing of my professional context, reading history, meeting notes, and founder relationship network.
So far, I have processed over 20 books using this method: Amplified (Dion Lim), The Autobiography of Bertrand Russell, Designing Your Life, The Tragedy of Child Prodigies, The 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 Chödrön), A Brief History of Everything (Ken Wilber), and so on.
Each book enriches this "mega-brain." The second mirror knows what the first mirror knew, and the twentieth mirror knows what the previous nineteen mirrors knew.
How the Book-Mirror Gets Better Through Iteration
The first time I did a book mirror, the result was disastrous.
In the initial version, there were three factual errors about my family. It stated that my parents were divorced, which was not true; it also said I grew up in Hong Kong, whereas I was actually born in Canada. These were foundational mistakes that could severely damage trust if shared.
So I added a mandatory fact-checking step. Now, before each mirror delivery, there is a cross-modal assessment of known facts in the mega-brain. Opus 4.71M catches accuracy errors; GPT-5.5 uncovers missing context; and DeepSeek V4-Pro evaluates if certain content sounds too generic.
Later on, I upgraded it to a deep retrieval based on GBrain tool invocation. The initial version was good at synthesis but weak in specificity. The third edition started conducting section-by-section brain searches. Each item in the right column references a real brain page.
When the book discusses how to handle difficult conversations, it doesn't just summarize some general principles. It pulls up real meeting notes between me and co-founders going through tough discussions; or a random idea that popped up during a Thursday chat with my brother James; or an instant messaging chat from when I was 19 with my college roommate. It feels surreal.
This is the practical meaning of "skillification" (or /skillify in GBrain). I refined my first manual attempt into a reproducible pattern, written as a tested skill file containing trigger conditions and edge cases. Subsequent fixes will continue to compound in all future book mirrors.
The Skill of Creating Skills
Here comes the true recursion, which I believe is one of the greatest insights.
The systems that support my daily life do not appear as a monolithic entity. They are assembled from individual skills. And these skills themselves are created by another skill.
Skillify is a "meta-skill" — a skill used to create new skills. Every time I encounter a workflow that will be repeated in the future, I say, "Skillify this." It then looks back at what just happened, extracts the reproducible pattern, writes it into a tested skill file with trigger conditions and edge cases, and registers it in the resolver.
The aforementioned book-mirror pipeline is skillified from the first time I manually completed this process. The meeting prep workflow is also the same: when I realize I am doing the same steps before each call, I skillify it.
Skills can be combined. Book-mirror calls brain-ops for storage, enrich for context addition, cross-modal-eval for quality assessment, and pdf-generation for output. Each skill focuses on one thing, but they can be strung together to form complex workflows.
When I improve one skill, all workflows using that skill automatically get better. There is no longer an issue of "I forgot to mention this edge case in the prompt." The skill remembers.
The Meeting Where I Did My Own Prep
Demis Hassabis came to YC for a fireside chat. Sebastian Mallaby's biography about him has just been published.
I had the system help me prepare.
In less than two minutes, it pulled out: Demis's full brain page — the content on this page has been accumulating from articles, podcast transcripts, and my own notes for months; his publicly expressed views on the AGI timeline, such as "50% through scaling, 50% through innovation," and his belief that AGI still needs 5 to 10 years; key points from Mallaby's biography; his explicitly mentioned research priorities, including continual learning, world modeling, and long-term memory; cross-references between his and my publicly discussed AI views; three demo scripts used to showcase this "brain's" multi-hop reasoning ability in conversation; and a set of dialogue entry points designed based on overlaps and divergences in our worldviews.
This is more than just a better Google search. It is a context-aware preparation: the system not only leverages the long accumulation of information about Demis but also takes into account my own stance and the strategic objectives of this conversation.
It prepares not just facts but perspectives.
What a 100,000-Page Brain Looks Like
I maintain a structured knowledge base of about 100,000 pages.
Everyone I encounter has a page that includes a timeline, a status bar—representing the current ground truth, ongoing threads, and a rating. Every meeting has a transcript, a structured summary, and a process I call "entity propagation": after each meeting, the system traverses every person and company mentioned in the meeting and updates their corresponding brain page with the discussion content.
Every book I read receives a chapter-by-chapter book mirror. Every article, podcast episode, and video I engage with gets ingested, tagged, and cross-referenced.
The schema is simple. Each page has three parts: at the top is the "compiled ground truth"—the current best understanding; at the bottom is a timeline that only appends and never modifies, sequentially documenting events; and on the side is the original data sidecar for source material.
You can think of it as a personal version of Wikipedia. Each page is continuously updated by an AI that attends meetings, reads emails, watches talks, and digests PDFs.
Here's an example to demonstrate how this system compounds.
I met a founder in office hours. The system would create or update his personal page, company page, cross-reference meeting notes, check if I've met him before—if so, it surfaces what we last discussed; it would check his application materials, pull in the latest metrics, and identify if anyone in my portfolio companies or contacts can help with the issue he's tackling.
By the time I walk into the next meeting with him, the system has prepared a full context package.
This is the difference between a "filing cabinet" and a "nervous system." A filing cabinet merely stores things; a nervous system connects them, tracks changes, and surfaces the most relevant information in the moment.
Architecture
Here’s how it works. I believe this is the right way to build personal AI, so I open-sourced everything, and you can set it up yourself.
The Harness is thin. OpenClaw is the runtime. It takes my messages, figures out which skill to apply, and then dispatches. It's only a few thousand lines of routing logic. It doesn’t know about books, meetings, or founders; it’s just responsible for routing.
The Skills are thick. There are now over 100 skills, each a self-contained markdown file with detailed instructions for a specific task. You've seen book-mirror and meeting-prep earlier. Here are a few more skills that come with GBrain:
meeting-ingestion: After each meeting, it pulls the transcript, generates a structured summary, then traverses every person and company mentioned in the meeting, updating their brain pages with the discussion. The meeting page itself isn’t the final product; the real value is in disseminating this information back to every individual and company page.
enrich: Give it a person's name. It pulls information from five different sources, consolidates everything into one brain page, including career trajectory, contact details, meeting history, and relationship context. Every assertion has a source reference.
media-ingest: Handles video, audio, PDFs, screenshots, and GitHub repos. It transcribes content, extracts entities, and archives the material in the right brain locations. I often use this for YouTube videos, podcasts, and voice memos.
perplexity-research: This is network research with brain-enhanced capabilities. It searches the web by Perplexity, but before synthesizing, it first checks what's already known in the brain to tell you what information is truly new and what you've already captured.
I’ve also built dozens of skills for my work, which I'll likely open-source: email-triage, investor-update-ingest—it identifies portfolio updates in my inbox and extracts metrics onto company pages; calendar-check—for detecting schedule conflicts and impossible travel arrangements; and a whole research stack I use for public affairs work.
Each skill encodes a type of operational knowledge that might take a few months for a new human assistant to learn. Some ask me how I "prompt" my AI. The answer is: I don’t prompt. The skill itself is the prompt.
The data is thick. The brain repo contains 100,000 pages of structured knowledge. Everyone I've met, every company, every meeting, every book, every article, every idea I've encountered is interconnected, searchable, and growing every day.
The code is also dense. The code feeding it is equally crucial: scripts for transcription, OCR, social media archiving, calendar syncing, API integrations. But where the true compounding value lies is in the data.
I run over 100 cron jobs every day, monitoring everything I follow: social media, Slack, email, and any other information I keep an eye on. My OpenClaw/Hermes Agents also watch these for me.
Models are interchangeable. For accuracy, I use Opus 4.7 1M; for recall and exhaustive extraction, I use GPT-5.5; for creative work and third-person perspective, I use DeepSeek V4-Pro; for speed, I use Groq with Llama. Skill determines which task calls which model. Harness doesn't care.
When someone asks "which AI model is the best," the answer is: you're asking the wrong question. The model is just the engine; everything else is the entire vehicle.
The 2 AM Builder and a Compound System
People ask me questions about productivity. But that's not how I think.
I think in terms of compounding.
Every meeting I attend adds to this brain. Every book I read enriches the context for the next. Every skill I build makes the next workflow faster. Every updated persona page sharpens the preparation for the next meeting.
The system today is ten times what it was two months ago. In another two months, it will be another ten times what it is now.
When I'm still coding at 2 AM—and yes, I do this often because AI has reignited the joy of building for me—I'm not just writing software. I'm adding capabilities to a system that gets better every hour.
100 cron jobs run around the clock. Meeting ingestion happens automatically. Email triage runs every 10 minutes. The knowledge graph enriches itself from every conversation. The system processes daily transcripts and extracts real-time patterns I may have missed.
This is not a writing tool, not a search engine, and not a chatbot.
This is a truly operational second brain. It's not a metaphor; it's a running system: 100,000 pages of content, over 100 skills, 15 cron jobs, and the context accumulated from every professional relationship, every meeting, every book, and every idea I've engaged with over the past year.
I've open-sourced the entire tech stack. GStack is a coding skills framework with over 87,000 stars, which I used to build this system. When the agent needs to write code, I still use it as a skill in the OpenClaw/Hermes Agent. There's also a great programmable browser that supports both headful and headless modes.
GBrain is the knowledge infrastructure. OpenClaw and Hermes Agent are harnesses—you can choose either one, but I usually use both. The data repository is also on GitHub.
The core insight is simple: the future belongs to individuals who can build compounding AI systems, not to those who only use enterprise-owned centralized AI tools.
The difference between the two is like the difference between writing a diary and having a nervous system.
How to Get Started
If you also want to build such a system:
Start by choosing a harness. You can use OpenClaw, Hermes Agent, or build one from scratch based on a Raspberry Pi. The key is to keep it lightweight. The harness is just a router. You can deploy it on a spare computer at home and access it via Tailscale; or you can host it on cloud services like Render or Railway.
Then, create a "brain" using GBrain. Initially inspired by Karpathy's LLM Wiki, I implemented it in OpenClaw and later expanded it into GBrain. It's the best retrieval system I've tested: achieving a 97.6% recall rate on LongMemEval and outperforming MemPalace in the retrieval phase without calling LLM. It comes with 39 installable skills, including all the content mentioned in this article. Installation is just a single command away. You'll get a git repo where every person, meeting, article, and idea has its own page.
Next, go do something genuinely interesting. Don't start by planning your skill tree. Instead, complete a specific task: write a report, research a person, download a season of NBA scores and develop a prediction model for your sports bets, analyze your investment portfolio, or do anything you truly care about. Use your agent to do it, iterate continuously until the results are good enough, and then run Skillify— the meta-skill mentioned earlier— to extract the patterns into a reusable skill. Then run check_resolvable to confirm that this new skill has been integrated into the resolver. This loop will turn one-off work into a reusable infrastructure that can compound over time.
Keep using it and examine the output carefully. This skill will initially be very ordinary. That's the point. Use it, read the content it generates, and when you find something off, run cross-modal eval: pass the output to multiple models to score each other based on dimensions you care about. This is how I initially discovered factual errors in the book-mirror. Fixes were written into the skill, making each mirror cleaner since then.
After six months, you'll have something no chatbot can replicate. Because the real value is not in the model itself, but in you teaching this system to understand your specific life, work, and way of reasoning.
The first thing I built with this system was terrible. By the one hundredth, it was a system I could trust with my calendar, inbox, meeting prep, and reading list. The system is learning, and so am I. The compounding curve is real.
Thick skill, thick code, thin harness. LLM itself is just an engine. You can entirely build your own car.
Everything I've described here— all the skills, the book mirror pipeline, cross-modal eval framework, skillify loop, resolver architecture, and over 30 installable skill packs— has been open-sourced and is freely available on GitHub.
Go build.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
