How to Make Claude Code Your AI Engineering Team

GStack is an open-source toolkit built by YC President & CEO Garry Tan that turns Claude Code into an AI engineering team — with skills for office hours, design, code review, QA, and browser testing.

Use it with Claude Code or Codex or Cursor. It’s free and open source: https://github.com/garrytan/gstack

In this video, Garry walks through how GStack works, starting with Office Hours, a skill modeled after real YC partner sessions that pressure-tests your idea before you write a line of code. He demos it live, going from idea through adversarial review, design mockups, and automated QA in a single session.

(Channel description, Y Combinator.)

A 21:49 demo + monologue by Garry Tan (YC President & CEO; engineer-by-background — Stanford CS, Palantir employee #10, Posterous co-founder sold-to-Twitter, built first version of YC’s Bookface). Published 23 April 2026 to the Y Combinator YouTube channel — the day before Diana Hu’s “How to Build a Company with AI from the Ground Up” on the same channel. Two YC-partner-vantage sources on AI-native-engineering in two consecutive days — the YC-batch-context anchor pair. ASR caption track; 15 chapters; no manual track.

TL;DR

Three substantive contributions:

1. GStack as an open-source harness-of-skills shipping at venture pace: Tan built GStack in 3 weeks and reports it “now has more GitHub stars than Ruby on Rails.” The toolkit wraps Claude Code (or Codex or Cursor) with named skills — Office Hours (16-YC-partners-distilled adversarial-review of startup ideas), Design Shotgun (multiple AI design variants in ~60 seconds), Code Review (staff-level bug-catching after plan), QA / Browser Testing (Playwright+Chromium wrapped as a CLI rather than via Claude-in-Chrome MCP), Ship (pre-merge gate). The wiki’s first founder-CEO-vantage open-source-harness anchor — a YC-president-of-a-Fortune-100-class-accelerator publicly shipping his own harness implementation as a worked example, not just describing one.

2. The AI-founder-type archetype operationalised at YC-president scale: Tan reports 10 to 15 parallel Claude Code sessions running simultaneously, sometimes 3-4 on the same project; “about 400 PRs to review right now”; “10, 15, 20, sometimes 50 PRs in any given day, depending on the number of meetings I have.” Direct operationalisation of Hu’s AI founder type archetype — “still builds, still coaches and leads by example. If you’re the founder, this needs to be you at the forefront.” Tan IS the worked example Hu describes.

3. Steve Yegge “Gas Town” stage 7 invoked explicitly: “There’s this idea of trying to get a level 8 software factory and GStack does not get you to level 8, but I do think it gets you to level seven.” The wiki’s second source explicitly invoking Yegge’s eight-stages-of-dev-evolution-to-AI framing (after Böckeler 2026). Two-source threshold met on Yegge’s framework as a named pattern in the wiki.

Plus: the ADHD-CEO-vs-autistic-CTO model-allocation framing — “by default [Claude Code] uses Claude. And I think Opus 4.6 is sort of ADHD CEO. He’s the guy you want to get a beer with and he’s got a billion ideas, but when the going gets tough, you got to call in your autistic CTO and that’s Codex.” First-party-CEO articulation of model-allocation-by-personality-fit-to-task as a working practice.

What was actually ingested

Full 21:49 transcript across 15 chapters. ASR transcript with some name-recognition ASR errors flagged below (“Palunteer” for Palantir; “Gary” for Garry; “GSAC” for GStack; “Gritan” for Garry Tan in the GitHub URL).

GStack architecture — a harness-of-skills for Claude Code

GStack Skill	What it does	Anchored against
Office Hours	Adversarial review of startup ideas — distillation of 16 YC partners × thousands of hours × many years; “10% strength version of what we do at YC every day”. Asks the question that determines everything else: “What’s the strongest evidence that you have that someone actually wants this?”	Maps onto agent-driven spec-elicitation candidate concept (now visible in: Rufus / Thompson / Böckeler / Tan — four-source convergence)
Plan / CEO Review / Adversarial Review	Multi-step adversarial review of a design doc; auto-fixes issues; tracks scored improvement (Tan’s demo: 6/10 → 8/10 in two rounds; “automatically caught and fixed 16 issues”)	Maps onto Böckeler’s enhanced-lint-messages-as-good-prompt-injection at the design-doc layer
Design Shotgun	Generates multiple AI design variants in ~60 seconds and presents them as choices to the developer	Direct operationalisation of Thompson’s “feel like Steve Jobs picking from nine designs” developer self-description
Code Review	Staff-level bug-catching after the plan ships code; “finding bugs that might not have been in the plan mode”	Maps onto Guthrie’s human-review + eval pattern at the harness-built-in layer
SLQA / SL browse (browser testing)	Wraps Playwright+Chromium at the CLI level; “now your Claude Code and any agent can actually just use the browser” — screenshots, complex interactions, click/fill, download media, regression tests, CSS-bug-assessment. “Claude in Chrome MCP is one of the worst pieces of software I’ve ever used” — Tan’s explicit motivation for the rewrite	Direct convergence with Böckeler’s CPU-deterministic feedback-harness; Playwright-as-CLI as the wiki’s first vendor-published browser-control-without-MCP worked example
Ship	Pre-merge gate; “the last step before to make sure that your PR is ready to land on main”	Maps onto Guthrie’s evals-as-CI-check

The collective architecture is the wiki’s first package-everything-in-one-toolkit worked example — distinct from Anthropic’s Skills (modules to load) or Lopopolo’s golden-principles GC (in-vendor pattern). GStack ships as a single open-source repo with all skills bundled. Hu’s harness templates future-prediction (“pick a workflow topology, instantiate a pre-built harness rather than rebuild”) is partially realised here at startup-development scope.

Multi-session orchestration via Conductor

Tan demonstrates GStack inside Conductor — a multi-session Claude Code orchestrator (Dangling first-mention; not previously in the wiki). Workflow:

“I run 10 to 15 parallel Claude Code sessions all at the same time. I might in one session be running office hours on a brand new idea. … I have multiple open-source projects with tens of thousands of stars. And I’m probably sitting on about 400 PRs to review right now.”

The Conductor + GStack + multi-session pattern is the worked example of Hu’s queryable-organization-with-agents-everywhere construct, but for a single-developer-with-many-projects rather than a company-with-many-functions. Same pattern, different scope.

Tan reports the bottleneck shift characteristic of multi-session work: “once the agent was doing all the work of planning and design and coding it, I found myself sitting there doing QA, probably the least fun part of software development. So that made it very, very important for me to try to automate that.” QA → SLQA / SL browse — the harness extends in the direction of the human’s remaining bottleneck.

Tan’s worked example — a tax-app idea

Tan demos the full workflow live with a tax-document-aggregator idea:

1. Office Hours asks: “What’s the strongest evidence that you have that someone actually wants this?”; “How many bank accounts do we have and which one sent a 1099?” — multi-step pressure-test.
2. The “this idea just got way bigger” inflection: Office Hours reframes the idea as “a funnel — the hook is we’ll find all your 1099s for you, the expansion is matchmaking and lead-gen for tax preparers. It’s a classic wedge strategy.” Document-aggregation could charge $2-5/year; transaction-percentage of tax-preparer matchmaking is “10× more” monetisable. Office Hours surfaces business-model expansion not just feature-list refinement.
3. Plan mode: Tan-as-CEO selects among three approaches presented by the agent (A: small Gmail-OAuth-only / B: full-stack Gmail + AI browser + CPA marketplace / C: CPA-first inverse go-to-market). Selects B with a refinement (“use the browser interaction to skip Google OAuth entirely”).
4. Adversarial Review: catches 16 issues across no-failure-handling / no-privacy-section / 2FA-handoff; auto-fixes most; score improves 6→8 of 10.
5. Design Shotgun: generates three design directions; presents as choices.
6. Code + Code Review + Ship.

The demo is fast — entire idea-to-near-shippable-design in ~10 minutes of video time. Compresses what Thompson 2026’s 75-developer survey reports anecdotally into a single-screen worked example at YC-partner scale.

Convergence and contradictions

Source	Connection
Y Combinator 2026 (24 April)	The YC pair: Tan (23 April) + Hu (24 April) — two consecutive-day YC-partner-vantage sources on AI-native engineering. Tan IS Hu’s AI founder type archetype operationalised: still builds, still ships, runs 10-15 parallel sessions at YC-president scale. Hu names the framework (closed-loop / queryable-org / software-factories / 1,000× engineer / token-maxing); Tan ships an open-source instantiation (GStack) of the framework’s first principles
Böckeler 2026 (Thoughtworks)	Yegge Gas Town two-source convergence. Böckeler invokes Yegge’s eight-stages-of-dev-evolution-to-AI as audience-orienting framing; Tan invokes the same essay with the specific “level 7 vs level 8” terminology Yegge defines. Direct two-source agreement on Yegge as the named-pattern source for the multi-parallel-session practice. Plus: Tan’s SLQA/SL-browse-wraps-Playwright is the specific implementation of Böckeler’s CPU-deterministic-feedback harness, with the Claude-in-Chrome-MCP-is-terrible anti-pattern named explicitly
Claude Code engineering)	Fung described what running an agent-native engineering team looks like inside the vendor; Tan shows what running an agent-native-engineering-of-one looks like at YC-president scale. Same dogfooding principle at different organisational scales
OpenAI Codex 2026	Lopopolo: in-vendor at OpenAI Codex (5 months / ~1M LOC / 0 manually-written lines). Tan: open-source-toolkit-shipping-in-3-weeks at YC. Two divergent strategies (in-house vendor harness vs open-source community harness) for the same problem
Thompson 2026 (NYT The Daily)	Thompson’s “feel like Steve Jobs picking from nine designs” developer self-description is literally what GStack’s Design Shotgun does — multiple AI design variants in ~60s as choices. Operator-side cultural observation → operator-side toolkit instantiation
LangChain ADLC 2026	GStack’s skill set maps onto Chase’s ADLC: Office Hours/Adversarial Review = Build-phase-spec-elicitation; Code Review = Test; SLQA/Browser = Test simulations; Ship = Deploy gate. GStack is Chase’s ADLC operationalised as a single CLI toolkit rather than as a vendor product portfolio
Braintrust 2025	Both name evals-as-CI / pre-merge-gate as a harness pattern; Tan ships it (Ship skill); Guthrie’s Braintrust platform offers the eval-platform layer GStack’s Ship skill could integrate with

Contradictions

None substantive. Tan’s framing is enthusiastic but operationally grounded.

Linked entities and concepts

Existing wiki entities reinforced: Y Combinator (Dangling from Hu 2026; now two-source — promotion candidate but defer entity-page creation to the cross-cutting work in this batch).

Concept pages updated:

• vibe-coding — Tan’s “the barrier to building just collapsed … what are you going to build?” peroration; live demo as worked example of the floor-raising thesis at solo-founder scale.
• agent-harness — GStack as the first founder-CEO open-source-toolkit harness instantiation.
• agentic-engineering — Tan’s 10-15-parallel-sessions / 400-PRs-in-review at YC-president scale as the AI-founder-type archetype.

Dangling (single-source first-mention, deferred per the author-entity-promotion rule):

• Garry Tan — YC President & CEO; engineer (Stanford CS); ex-Palantir #10; Posterous co-founder; built Bookface.
• GStack — Garry Tan’s open-source toolkit (github.com/garrytan/gstack). Strong promotion candidate — likely to recur on future Claude-Code-skill-ecosystem ingests.
• Conductor — multi-session Claude Code orchestrator Tan demos GStack inside of.
• Palantir — Tan’s first decade employer; ASR rendered as “Palunteer”.
• Posterous — Tan’s pre-YC co-founded micro-blogging startup (sold to Twitter).
• Bookface — YC’s internal social platform; Tan built v1.

Concept candidates surfaced (not yet promoted):

• Model-allocation-by-personality-fit-to-task — Tan’s “ADHD CEO Claude Opus 4.6 vs autistic CTO Codex” framing. Anthropomorphic but operationally meaningful. Single-source.
• Open-source-toolkit-of-skills-shipping-in-weeks — GStack as worked example. Single-source as a pattern-named-as-such.
• Bottleneck-shift-to-QA-when-coding-is-automated — Tan: “once the agent was doing all the work … I found myself sitting there doing QA, probably the least fun part.” Single-source as a named-bottleneck.

Open questions raised by this source

• GStack repository deep-dive — would substantiate the skill architecture claims. Primary-source target: github.com/garrytan/gstack.
• Conductor as a separate tool — Tan uses it as the multi-session shell. First-party documentation on Conductor’s architecture and relationship to Claude Code would let the wiki ingest it as a distinct entity.
• “More GitHub stars than Ruby on Rails in 3 weeks” — staggering velocity claim. Worth tracking the 3-month / 6-month adoption arc.
• Whether the GStack architecture transfers to non-individual-developer settings — Tan operates at YC-president-with-many-side-projects scale; would the same skills work for a 100-developer org? Open empirical question.