the deck
AI workflow · Garry Tan · via The AI Corner

Personal AGI: own your context, rent the models

Garry Tan (YC president) claims a ~400× output multiplier from an 80,000-page personal knowledge system. The point isn't a smarter model — it's that your context is an asset you build, while the model is a commodity you rent.

Written by Ruben Dominguez · The AI Corner · 31 Aug 2026 — on Garry Tan's "personal AGI" · reformatted for internal use
Idea I

One you consume. One you build.

Frontier models are a product you consume. Your context is an asset you build.

Every lab ships a better model every few months, and everyone gets it at once — so the model itself is a commodity. What isn't commodity is the accumulated context around your work: your history, your decisions, your conventions, your procedures. Rent the model; own that.

same frontier models rented — resets each release owned — compounds daily
Same models feed both curves. The difference is what you keep.

What this section means

If your AI setup would reset to zero when you switch vendors, you've been renting everything. The moat is the part that survives the switch.

Idea II

What "personal AGI" actually is

Not a model. A private, plain-text repository — usually markdown — in two halves:

  • MMemory — facts, decisions, project history, people, your conventions. It accumulates.
  • SSkills — reusable step-by-step procedures the agent can execute ("how I design an eval", "how I run a release"). Judgment, written down once.

It gets sharper every day, because it's fed your specific history — not every few months when a vendor ships. Tan reports an 80,000+ page system and a claimed ~400× multiplier.

YOUR AGI memory/ skills/ output
Two folders and an agent. That's the whole architecture.
Idea III

The gap is workflow discipline, not model quality

Everyone has the same frontier model. Yet some people get 2× out of it and some get 100×. That spread is process — capture discipline, skill files, a clean division of labour — not access to a better model.

2× 100× same models · the gap is workflow discipline
25% of YC's W25 batch shipped 95%-AI codebases — and grew fastest.

What this section means

Stop waiting for the next model to change your output. The lever you actually control is the workflow around it.

Idea IV

Write the skill file before you'd hire for it

Any task you repeat is a candidate. If you'd consider hiring someone to do a thing, you already know the procedure well enough to write it down — so write it as a skill file the agent can run. Repeated task → reusable skill. The library becomes your leverage.

What this section means

Every annoying task you solve twice without writing it down is leverage you threw away.

Idea V

Separate judgment from computation

You decide what matters and why. The model does the volume work. Keep that line clean and you scale without handing over the part that's actually you.

youjudgment modelcomputation what & why → ← result to check
Hand over the volume, keep the direction.
Idea VI

Keep control of the repository

Plain text. Version-controlled. Portable. If your memory and skills live inside one vendor's product, you don't own them — and the compounding stops the day you leave. Markdown + git is the whole requirement.

The five principles

1 · own context, rent models  ·  2 · skill files precede hiring  ·  3 · separate judgment from computation  ·  4 · repeated task → reusable skill  ·  5 · retain control of your repo.

The tools already exist

Nothing here is waiting on a better model. Ownership and discipline are what separate compounding growth from marginal gains.

Ruben Dominguez · The AI Corner · on Garry Tan's "personal AGI"
Apply it · 16 years in tech

16 years of judgment, most of it unwritten

You already have the context Tan is talking about — it's in your head, not in a repo. Structuring it into memory + skills turns a decade and a half of instinct into a compounding, portable, demonstrable asset.

Idea I → applied

Own the part that survives a job change

Your value right now is largely trapped in one employer's context. A private repo of how you do AI engineering is the version of that which moves with you — into a frontier lab, a startup, or your own thing.

The move

Start brain/ today: memory/, skills/, log/, one MEMORY.md index, git init.

Idea II + IV → applied

Turn your judgment into skill files

  • 1eval-design.md — how you decide what to measure and why.
  • 2pr-review-for-safety.md, rag-regression-debug.md, incident-runbook.md, paper-repro-checklist.md.
  • 3Rule: any task you did 3× this month becomes a skill file this week.

Why it matters for the lab goal

These files are judgment on display — exactly what a frontier lab screens for, and far more distinctive than another tutorial.

Idea III → applied

Use the workflow to ship the flagship solo

The eval suite / paper repro / OSS contribution you keep meaning to build — the personal-AGI workflow is what lets one person produce lab-quality work at pace. You direct; the model executes against your skill files.

Idea V → applied

Narrate the judgment / computation split

In every public writeup, show the line: "I decided this benchmark matters because X; the harness was generated." That division is the senior skill — make it visible.

Renting everything

  • Prompts live in a chat history you can't export
  • Judgment re-improvised every time
  • Nothing to show a hiring manager but a CV
  • Output tracks whatever model shipped last

Owning the context

  • memory + skills in git, vendor-agnostic
  • Judgment written once, reused and refined
  • A public playbook that reads as seniority
  • Output compounds regardless of release cycle

Write down what you already know

Sixteen years of judgment is the asset. It only compounds once it leaves your head and enters a repo you own.

framework: Garry Tan / Ruben Dominguez · application: yours to run
Apply it · starting fresh

Start the repo before you feel qualified to

You don't need 16 years to own context. Every project you debug and every setup you figure out is a skill file waiting to be written — and that repo doubles as your portfolio and your learning engine.

Idea I → applied

Own it in git, not in a chatbot

If your "AI workflow" is a pile of chat threads, you own nothing. Plain markdown + git from day one. Cost: five minutes. Payoff: it compounds instead of scrolling away.

The move

mkdir brain/{memory,skills,log} · git init · one MEMORY.md.

Idea II + IV → applied

Every "how do I…" you solve becomes a skill

  • 1set-up-a-python-project.md, debug-an-api-error.md, ship-a-small-app.md.
  • 2Beginner procedures are still procedures. Writing them is how you learn them properly.
  • 3A sanitized subset → blog posts. That's the "learn in public" trail hiring managers want.
Idea III → applied

The gap you can close now is discipline

You can't out-model anyone — but capture discipline is free. Write the note at the moment of friction. Most people don't, which is exactly why 2× and 100× users run the same model.

Idea V → applied

You keep the judgment even as a beginner

Decide what to build and why; let the model handle the boilerplate. Practising that split now is the habit that scales later.

Tutorial-hell workflow

  • Answers vanish into chat history
  • Re-Googling the same fix monthly
  • Nothing to show but "I'm learning"
  • Waiting for a better model to help

Personal-AGI workflow

  • Every fix → a skill file in git
  • Solve once, reuse forever
  • A growing public playbook + projects
  • Compounding from the model you already have

Capture at the moment of friction

Own the repo, write the skill, keep the judgment. The compounding starts on day one.

framework: Garry Tan / Ruben Dominguez · application: yours to run