Garry Tan recently published a long-form post, 《Meta-Meta-Prompting: The Secret to Making AI Agents Work》, detailing exactly how he built a suite of an AI agent–driven “second brain system.” He said that over the past five months, AI has already brought him back to being a builder—and even fundamentally changed how he goes about his day-to-day work as the CEO of Y Combinator.
YC CEO: The future belongs to people building compound AI systems
Garry Tan believes most people still treat AI as a chat window, but the real opportunity is to treat AI as an “operating system”: a system that can continuously remember, accumulate, update, and connect personal knowledge, meetings, reading, and workflows. He even said bluntly: “The future belongs to those who build compound AI systems, not those who only use large company centralized AI tools.”
(YC partner shares how to use AI from scratch to build a company; startups should treat AI as an operating system, not a tool)
This also aligns with what we reported earlier. Whether it’s the Y Combinator Summer 2026 Requests for Startups (RFS) or what YC partner Diana mentioned in Startup School, both point out that AI entrepreneurship is shifting from “improving individual productivity” to “rebuilding organizational and industry processes.” AI shouldn’t just be an efficiency tool companies occasionally use; it should be designed from day one to be the operating system for the entire company.
(YC releases 15 startup directions it wants to invest in for Summer 2026: AI entrepreneurship isn’t about stuffing a chatbot into a product)
AI helps him turn Buddhist books into a “life mirror”
Garry Tan shared that while reading Buddhist writer Pema Chödrön’s 《When Things Fall Apart》 recently, he first truly became aware of the power of personal AI. He had his AI system run a “book mirror” process: first break down the entire book’s 22 chapters, then have multiple sub-agents simultaneously do two things—summarize the author’s viewpoints and map each viewpoint to Garry Tan’s own life.
And it’s not generic spiritual comfort—it directly combines:
family background
entrepreneurship history
YC work
late-night notes
reading records
therapist discussion content
conversations with founders
The final output is a long “brain page” totaling 30,000 words.
For example, when the book discusses groundlessness, the system links it to a specific conversation he had with a founder the week before; when it discusses fear, it cites the behavioral patterns pointed out by the therapist; and when it discusses letting go, it connects to the sense of creative freedom he wrote about late at night.
Garry Tan said the entire process takes about 40 minutes. He believes even a therapist charging $300 per hour can’t complete a similar analysis within 40 hours, because humans can’t load all work context at the same time—reading history, meeting notes, interpersonal relationship maps—but AI can.
The real key isn’t the model—it’s the “skills system”
However, Garry Tan believes the truly important part of an AI agent isn’t a single model, but “skills.” His current system includes more than 100 AI skills and about 100,000 pages of a knowledge base.
He calls this kind of architecture:
Fat skills. Fat code. Thin harness.
Meaning:
The harness (runtime/router) should be thin
Skills should be fat
Real value lies in knowledge, workflows, and data
He currently uses Anthropic Claude Opus 4.7 for precision, GPT-5.5 for recall and extraction, DeepSeek V4-Pro for creative work, Groq + Llama for fast reasoning, while OpenClaw and Hermes Agent handle routing.
Garry Tan believes: “A model is just the engine, and everything else is the car.” AI agents are no longer just prompts; they’re compounding workflows. Garry Tan emphasized that he now almost never prompts AI. What matters is the skill system.
For example:
meeting-ingestion
media-ingest
enrich
perplexity-research
investor-update-ingest
email-triage
calendar-check
Each skill is a reusable, testable, composable workflow module. And most importantly, he even built a meta-skill called “Skillify.” When he finds that a certain workflow keeps repeating, he just inputs: skillify this
The system then analyzes what was just done, extracts repeatable patterns, creates a skill file, adds it to the resolver routing system, and accumulates it for all future workflows.
A 100,000-page knowledge base: AI starts to act like a neural system, not a filing cabinet
Garry Tan said he currently maintains a structured knowledge base of about 100,000 pages. Every person, company, meeting, book, Podcast, article, and idea gets its own dedicated page. And after each meeting ends, the AI automatically generates a transcript, creates summaries, updates the person pages, updates the company pages, updates the timeline, updates open threads, and updates relationship context.
This means AI is no longer just storing data—it’s starting to act like a “neural system.” He described a filing cabinet as merely storing things, while a neural system connects, reminds, updates, and derives conclusions.
The most important thing in the AI era is a personal compound system
In the end, Garry Tan’s core view is very clear: the strongest people in the future may not necessarily be those using the strongest models, but those who can build:
their own knowledge graph
their own workflows
their own skill system
their own personal AI OS
Because as every book, every meeting, every skill improvement, and every data update continuously accumulates, the entire AI system starts to show compounding effects. He even said that even though he’s still coding every day at 2:00 a.m., it’s not because there’s too much work—it’s because: “AI has given the joy of being a builder back to me.”
This article by Garry Tan: I now rarely give prompts to AI! YC CEO explains “compounding AI workflows” first appeared on Chain News ABMedia.
