Editor's Note: While most people still see AI as a smarter chat window, Y Combinator's current CEO, Garry Tan, is already trying to turn it into a personal operating system.

The underlying structure of personal productivity in the AI era is changing: models are just engines; what truly generates compound interest is the entire system built around personal knowledge, workflows, context, and judgment.

In this system, every meeting, every book, every email, and every connection is no longer an isolated piece of information but is continuously written into a structured 'second brain.' Every recurring task no longer relies on temporary prompts but is abstracted into reusable skills that are continuously iterated in subsequent work. In other words, AI doesn't just help people complete tasks; it helps individuals productize, systematize, and infrastructure-ize their own way of working.

Even more noteworthy is that the author proposes a personal path different from centralized AI tools: future competitiveness may not belong only to those who can use AI, but to those who can train a compound-interest AI system around their real life and work. Chatbots give answers, search engines provide information, but a true personal AI system remembers your background, understands your context, inherits your judgment, and becomes stronger with every use.

This is also the most enlightening part of this article: the value of AI does not lie in what it generates once, but in whether it can become a nervous system that continuously accumulates, connects, and improves. For individuals, this is perhaps the true starting point of an 'AI-native way of working.'

Below is the original text:

People always ask me why I spend my nights coding until 2 a.m. I have a job, and a heavy one—I am the CEO of Y Combinator. We help thousands of entrepreneurs every year achieve their dreams: starting real, revenue-generating, fast-growing startups.

Over the past 5 months, AI has turned me back into a builder. By the end of last year, the tools were good enough for me to start building again. Not toy projects, but systems that can truly compound. I want to show you with concrete examples what it actually looks like when you stop treating your personal AI as a chat window and start treating it as an operating system. I'm open-sourcing this stuff and writing about it 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' talks about the intelligent routing table; 'The LOC Controversy' discusses how every technologist can amplify themselves 100x to 1000x; 'Naked models are stupider' argues that models are just engines, not the whole car; and the 'skillify manifesto' explains why LangChain raised $160 million but gave you a squat rack and dumbbells without a training plan, while this article gives you the training plan you actually need.

That Book That 'Read Me Backwards'

Last month, I was reading Pema Chödrön's 'When Things Fall Apart.' The book is 162 pages, 22 chapters, about Buddhism's view on suffering, groundlessness, and letting go. A friend recommended it to me during a difficult time.

I had my AI do a 'book mirror.'

Specifically, this means: the system extracted the full content of all 22 chapters, then ran a sub-agent for each chapter, doing two things simultaneously: summarizing the author's ideas and mapping every point to my real life.

Not vague platitudes like 'this also applies to leaders,' but very specific mappings. It knows my family background: immigrant parents, father from Hong Kong and Singapore, mother from Myanmar. It knows my professional context: I'm managing YC, building open-source tools, mentoring thousands of founders. It knows what I've been reading recently, what I'm thinking at 2 a.m., what issues I'm working on with my therapist.

The final output was a 30,000-word 'brain page.' Each chapter was presented in two columns: one column for what Pema was saying, the other for how that content mapped to what I'm 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 behavioral patterns my therapist had pointed out; 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. A therapist charging $300 an hour couldn't do this in 40 hours, even after reading the book and applying it to my life. Because they don't have my full professional context, reading history, meeting notes, and founder network loaded and cross-referenceable.

So far, I've processed over 20 books this way: 'Amplified' (Dion Lim), 'The Autobiography of Bertrand Russell,' 'Designing Your Life,' 'The 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 Chödrön), 'A Brief History of Everything' (Ken Wilber), etc.

Each book makes this 'brain' richer. The second mirror knows the content of the first, the twentieth mirror knows all the content of the previous nineteen.

How Book-Mirror Got Better Through Iteration

The first time I did a book mirror, it was terrible.

In the first version, there were three factual errors about my family. It said my parents were divorced, but they're not; it said I grew up in Hong Kong, but I was actually born in Canada. These were basic mistakes that would have destroyed trust if I shared the results.

So I added a mandatory fact-checking step. Now, every mirror runs a cross-modal evaluation against known facts in the brain before delivery. Opus 4.7 1M catches precision errors; GPT-5.5 finds missing context; DeepSeek V4-Pro judges if something sounds too generic.

Later, I upgraded it to deep retrieval based on GBrain tool calls. The initial version was good at synthesis but weak on specificity. The third version started doing section-by-section brain searches. Every item in the right column would cite a real, existing brain page.

When the book talked about handling difficult conversations, it wouldn't just summarize generic principles. It would pull up real meeting notes from my sessions with founders who were having tough conversations with co-founders; or an idea that popped up during a casual chat with my brother James on a Thursday; or an instant messenger chat record from when I was 19 with my college roommate. It feels surreal.

This is what 'skillification' (/skillify in GBrain) means in practice. I distilled that first manual attempt into a repeatable pattern, wrote it into a tested skill file with triggers and edge cases. Since then, every fix compounds in all future book mirrors.

The Skill That Can Create Skills

Here's where it gets truly recursive, and I think this is the biggest insight.

The system that powers my daily life didn't appear as one giant monolith. It was assembled from skills. And those skills themselves were created by another skill.

Skillify is a 'meta-skill'—a skill for creating new skills. Whenever I encounter a workflow I'll repeat in the future, I say: 'Skillify this.' It then looks back at what just happened, extracts the repeatable pattern, writes it into a tested skill file with triggers and edge cases, and registers it with the resolver.

The book-mirror pipeline I mentioned earlier was skillified after I did it manually the first time. The meeting-prep workflow was the same: when I realized I was doing the same steps before every call, I skillified it.

Skills can be composed. Book-mirror calls brain-ops for storage, enrich for context supplementation, cross-modal-eval for quality assessment, pdf-generation for output. Each skill focuses on one thing, but they can chain together to form complex workflows.

When I improve one skill, all workflows using that skill automatically get better. No more 'I forgot to mention this edge case in the prompt.' The skill remembers.

The Meeting That Prepared Itself

Demis Hassabis came to YC for a fireside chat. Sebastian Mallaby's biography of him had just been published.

I had the system help me prepare.

In under two minutes, it pulled up: Demis's complete brain page—accumulated for months from articles, podcast transcripts, and my own notes; his publicly stated views on AGI timelines, like '50% scaling, 50% innovation,' and his belief that AGI is 5–10 years away; highlights from Mallaby's biography; his stated research priorities, including continual learning, world models, and long-term memory; cross-references between his publicly discussed AI views and mine; three demo scripts for showing off this 'brain's' multi-hop reasoning during the talk; and a set of conversation entry points based on where our worldviews overlap and diverge.

This wasn't just a better Google search. It was contextual preparation: the system used not only my long-accumulated information about Demis but also my own positions and the strategic goals of this conversation.

It prepared not just facts, but angles.

What a 100,000-Page Brain Looks Like

I maintain a structured knowledge base of about 100,000 pages.

Everyone I encounter gets a page with a timeline, a status bar—the current reality, open threads, and a score. Every meeting gets a transcript, a structured summary, and a process I call 'entity propagation': after each meeting, the system traverses every person and company mentioned and updates their brain page with the discussion content.

Every book I read gets a chapter-by-chapter book mirror. Every article, podcast, video I engage with is ingested, tagged, and cross-referenced.

The schema is simple. Each page has three parts: at the top is the 'compiled truth'—the current best understanding; below is an append-only timeline of events in chronological order; on the side is a raw data sidecar for source materials.

Think of it as a personal Wikipedia. Each page is continuously updated by an AI that attended the meeting, read the email, watched the talk, and digested the PDF.

Here's an example of how such a system compounds.

I see a founder during office hours. The system creates or updates their personal page, company page, cross-references meeting notes, checks if I've met them before—if so, surfaces what we talked about last time; it checks their application, pulls latest metrics, and identifies anyone in my portfolio or network who could help with their problem.

By the next time I walk into a meeting with them, the system has prepared a full context pack.

This is the difference between a 'filing cabinet' and a 'nervous system.' A filing cabinet just stores things; a nervous system connects them, flags what changed, and surfaces what's most relevant in the moment.

Architecture

Here's how it works. I think this is the right path to building personal AI, so I open-sourced the whole thing. You can build it yourself.

The harness is thin. OpenClaw is the runtime. It receives my messages, decides which skill applies, and dispatches. Only a few thousand lines of routing logic. It doesn't know about books, meetings, or founders; it just routes.

Skills are fat. There are over 100 now, each a self-contained markdown file with detailed instructions for a specific task. You've seen book-mirror and meeting-prep already. Here are a few others that come with GBrain:

meeting-ingestion: After each meeting, it pulls the transcript, generates a structured summary, then traverses every person and company mentioned, updating their brain page with the discussion. The meeting page itself isn't the end product; the real value is propagating that information back to individual and company pages.

enrich: Give it a person's name. It pulls information from five different sources, merges everything into a brain page, including career trajectory, contact info, meeting history, and relationship context. Every judgment has a source citation.

media-ingest: Handles video, audio, PDF, screenshots, GitHub repos. It transcribes, extracts entities, and files materials into the correct brain location. I use it often for YouTube videos, podcasts, and voice memos.

perplexity-research: This is web research with brain augmentation. It searches the web via Perplexity, but before synthesizing, it checks what the brain already knows, telling you what information is truly new versus what you've already captured.

I've built dozens more skills for my own work that I'll likely open-source later: email-triage, investor-update-ingest—which identifies portfolio updates in my inbox and extracts metrics to company pages; calendar-check—for detecting schedule conflicts and impossible travel; and a whole news research stack I use for public affairs work.

Each skill encodes operational knowledge that would take a new human assistant months to learn. People ask me how I 'prompt' my AI. The answer: I don't. The skill *is* the prompt.

Data is fat. The brain repo has 100,000 pages of structured knowledge. Every person, company, meeting, book, article, idea I've engaged with is connected, searchable, and growing daily.

Code is also fat. The code that feeds it matters too: scripts for transcription, OCR, social media archiving, calendar syncing, API integrations. But where the compound value truly sediments is in the data.

I run over 100 cron jobs daily checking everything I care about: social media, Slack, email, and any other signal I watch. My OpenClaw/Hermes Agents also watch these things for me.

Models are swappable. For precision, I use Opus 4.7 1M; for recall and exhaustive extraction, GPT-5.5; for creative work and third-person perspective, DeepSeek V4-Pro; for speed, Groq with Llama. The skill decides which model to call for which task. The harness doesn't care.

When people ask 'which AI model is best?' the answer is: you're asking the wrong question. Models are just engines; everything else is the car.

The 2 A.M. Builder, and a System That Compounds

People ask me about productivity. But that's not how I think.

I think about compound interest.

Every meeting I attend adds to this brain. Every book I read enriches the context for the next one. Every skill I build makes the next workflow faster. Every person page I update makes the next meeting preparation sharper.

The system today is 10x what it was two months ago. In another two months, it will be 10x again.

When I'm coding at 2 a.m.—and I often am, because AI has given me back the joy of building—I'm not just writing software. I'm adding capability to a system that gets better every hour.

100 cronjobs run 24/7. 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 patterns I didn't notice in real-time.

This isn't a writing tool, a search engine, or a chatbot.

It's a truly runnable second brain. It's not a metaphor; it's a running system: 100,000 pages, over 100 skills, 15 cron jobs, and the context accumulated from every professional relationship, meeting, book, and idea I've engaged with over the past year.

I've open-sourced the whole tech stack. GStack is a coding skill framework with over 87,000 stars, and I built this system with it. When an agent needs to write code, I still use it as a skill within my OpenClaw/Hermes Agents. It also has a great programmable browser, both headed and headless.

GBrain is the knowledge infrastructure. OpenClaw and Hermes Agent are harnesses—you can pick one, but I typically use both. The data repos are also on GitHub.

The core thesis is simple: the future belongs to individuals who can build compounding AI systems, not to those who only use corporate-owned, centralized AI tools.

The difference between the two is like the difference between keeping a diary and having a nervous system.

How to Start

If you also want to build such a system:

First, pick a harness. You can use OpenClaw, Hermes Agent, or build from scratch based on Pi. The key is to keep it light. The harness is just a router. You can deploy it on a spare computer at home and access it via Tailscale, or put it on a cloud service like Render or Railway.

Then, build a 'brain' with GBrain. I was initially inspired by Karpathy's LLM Wiki, implemented it in OpenClaw, and later expanded it into GBrain. It's the best retrieval system I've tested: 97.6% recall on LongMemEval, surpassing MemPalace in the retrieval step without calling an LLM. It comes with 39 installable skills, including everything mentioned in this article. Just one command to install. You get a git repo where every person, meeting, article, idea gets its own page.

Next, do one thing that's actually interesting. Don't start by planning your skill architecture. First, complete a concrete task: write a report, research a person, download a season of NBA scores and build a prediction model for your sports betting, analyze your portfolio, or do anything you genuinely care about. Do it with your agent, iterate until the results are good enough, then run Skillify—the meta-skill mentioned earlier—to extract the pattern into a reusable skill. Then run check_resolvable to confirm the new skill is hooked into the resolver. This cycle turns one-off work into infrastructure that keeps compounding.

Keep using it and read the output carefully. The skill will be mediocre at first. That's the point. Use it, read what it generates, and when you find something wrong, run cross-modal eval: give the output to multiple models and have them score each other based on the dimensions you care about. That's how I found the factual errors in book-mirror initially. The fix was written into the skill, and every mirror since has been cleaner.

Six months from now, you'll have something no chatbot can replicate. Because the real value isn't in the model itself, but in you teaching this system to understand your specific life, work, and judgment.

The first thing I made with this system was terrible. By the hundredth, it was a system I'd trust with my calendar, inbox, meeting prep, and reading list. The system is learning, and I'm learning. The compound curve is real.

Fat skills, fat code, thin harness. The LLM itself is just an engine. You can absolutely build your own car.

Everything I described here—all the skills, book-mirror pipeline, cross-modal eval framework, skillify loop, resolver architecture, and over 30 installable skillpacks—is already open-sourced and free on GitHub.

Go build.