Y CombinatorTokenmaxxing: How Top Builders Use AI To Do The Work Of 400 Engineers
CHAPTERS
- 0:00 – 0:47
Control vs. dependency: AI tools as Ferraris you must maintain
Garry frames the central question of the episode: whether you control your AI tools or they control you. He likens tools like OpenClaw to driving a Ferrari—extraordinary power, but brittle enough that you need the skills to debug and repair it when it fails.
- •Core question: control your tools vs. tools controlling you
- •OpenClaw/agents feel exhilaratingly capable, beyond what you'd expect from software
- •Brittleness is real: you still need to be a 'mechanic' who can fix failures
- •This moment feels historically significant for computing and builders
- 0:47 – 1:56
Returning to coding after 13 years—and why the output looks unbelievable
The hosts set up why this is notable: after years focused on investing/executive work, Garry is suddenly shipping massive amounts of code and open source projects while running YC. Garry himself is surprised by the scale, describing a jump to “400x” productivity by directing agents rather than typing everything manually.
- •Context: multi-year hiatus from hands-on coding
- •Recent burst: huge code output and fast shipping alongside a demanding day job
- •Claimed productivity leap: 400x vs. past baseline
- •Key shift: directing multiple agents rather than writing everything yourself
- 1:56 – 5:50
Garry’s List: rebuilding a startup-grade platform with Claude Code in days
Garry explains the real-world motivation behind building Garry’s List and how it sparked his return to coding. He recounts rebuilding a full-featured blogging platform (his earlier startup, Postrous/Posthaven lineage) dramatically faster and cheaper using Claude Code—adding modern RAG and agentic retrieval on top.
- •Civic motivation (e.g., education/algebra access) drives the project
- •Background: Postrous (blogs-by-email) → acquisition → Posthaven rewrite(s)
- •Third rebuild: ~5 days and ~$200 vs. prior multi-person, multi-month efforts
- •Adds RAG/agentic retrieval to support deep research and publishing workflows
- 5:50 – 6:17
Software that thinks like an investigative journalist (agentic research + sourcing)
The conversation shifts from “a blogging tool” to “software that does journalism.” Garry describes building workflows that retrieve, cross-reference, and cite many sources, using APIs and ‘boil the ocean’ research to generate deeply sourced reports and quotables.
- •Garry’s List does the work of investigative journalism, not just publishing
- •Pulls broad context: web search, Perplexity, X, Groq, and other sources
- •Cross-references competing narratives across many sources
- •Goal: more reality-representative output through exhaustive retrieval and citations
- 6:17 – 10:07
The rise of tokenmaxxing: 'boil the ocean' and pay for completeness
Garry introduces “tokenmaxxing” as a strategy: spend more tokens to get meaningfully better completeness, quality, and confidence. He argues knowledge work can be tokenmaxxed, but humans must still provide agency, values, and direction.
- •Tokenmaxxing = deliberately spending tokens to improve completeness and rigor
- •‘Boil the ocean’ mindset: don’t settle for minimal context or few sources
- •Applies beyond writing to coding and broader knowledge work
- •Humans supply intent/agency; machines supply scalable execution
- 10:07 – 14:16
Accidental creation of GStack: turning repeated prompts into reusable skills
Garry explains how GStack emerged unintentionally from repeated workflows while building Garry’s List. He found that forcing structured planning—like ASCII diagrams of data flows and state machines—reduced confusion, improved completeness, and became a reusable ‘skill’ system that went viral.
- •GStack started as personal prompt snippets in Apple Notes
- •ASCII diagrams (data flows, user flows, error cases) improve model performance
- •From ad-hoc prompts to structured skills: plan/review sections emerge
- •Viral spread of the early ‘plan-eng-review’ and evolution into CEO Plan
- 14:16 – 16:48
The workflow behind 400x output: queued PRs, plan→approve execution, and automation
Garry walks through his day-to-day workflow in Conductor: generating plans with skills, approving execution, and batching many PRs. The bottleneck became manual QA, which pushed him to automate browsing/testing—leading to a CLI daemon and QA agent behavior.
- •Workflow: plan mode → structured reviews → approve → agent executes
- •High-throughput batching: multiple PRs and feature queues
- •Manual testing/QA becomes the new constraint
- •Motivation to wrap Playwright and create automated browser-driven QA
- 16:48 – 20:59
Multi-agent orchestration: Claude Code + Codex as complementary minds
The episode details how Garry routes work between Claude Code and Codex, using each where it’s strongest. Claude acts as a product/CEO-like driver, while Codex is invoked for deep bug-finding and higher-difficulty technical scrutiny, with feedback looped back into the main agent flow.
- •GStack includes roles: CEO, design, dev-ex, ENG review, etc.
- •Claude Code strengths: speed, iteration, ‘ADHD CEO’ friendliness
- •Codex strengths: deeper technical critique and bug discovery
- •Workflow: Office Hours → CEO review → design/dev-ex → ENG review → Codex pass
- 20:59 – 25:52
Thin Harness, Fat Skills: where to use markdown vs. deterministic code
Garry formalizes the philosophy: don’t rebuild harnesses; invest in high-quality ‘skills’ (prompted procedures) that capture judgment and edge cases in plain language. Deterministic operations should live in code/tool calls, while ambiguous, contextual reasoning belongs in LLM instructions—backed by strong testing.
- •Harness = core loop/tooling; skills = reusable procedural knowledge in markdown
- •Analogy: wedding-planner checklist (skills) vs. Twilio calls (deterministic tools)
- •Common failure: encoding human judgment in brittle code instead of skills
- •Engineering discipline still matters: aim ~80–90% tests to avoid 'slop'
- 25:52 – 34:16
AI agents are like kit-car Ferraris: powerful, fragile, and fixable with other agents
Garry and Harj expand the Ferrari metaphor into an engineering reality: tools break, but the new trick is having agents fix agents. They trace the UX evolution from Stack Overflow to chat to agentic execution, arguing brittleness matters less when repair can be automated and continuously assisted.
- •OpenClaw/agent stacks can brick themselves; reliability is still maturing
- •Progression: Stack Overflow → ChatGPT → Claude Code (execute) → OpenClaw (agentic)
- •Meta-solution: keep an agent available to repair other agents/tools
- •Expectation: today’s fragility will feel normal and improve rapidly
- 34:16 – 35:53
The future of personal AI: owning your data, prompts, and integrations
Garry predicts that everyone will have a personal AI, and that the key choice is ownership and control. He contrasts personal, user-owned systems with corporate-controlled feeds and opaque incentives—calling this a PC-revolution-scale shift.
- •Prediction: personal AI becomes universal
- •Critical distinction: user-owned data/integrations/prompts vs. corporate-controlled systems
- •Opaque algorithms resemble social feeds; incentives may not align with users
- •Defining question returns: do you control your tools or not?
- 35:53 – 41:29
Buying back time with tokens: why token spend is like SF rent for founders
The hosts confront the cost objection: the best capabilities often require premium models and high token burn. Harj compares tokenmaxxing to paying high rent in SF for leverage and serendipity—expensive upfront but costly not to do if it accelerates outcomes and learning.
- •Gap in experience: free/basic tiers vs. latest models + high context windows
- •Tokenmaxxing reframed as an investment, not a cost center
- •Analogy: SF rent—expensive, but 'more expensive not to' for ambitious outcomes
- •Tokens buy ‘machine time’: leverage to reclaim scarce founder/creator time
