Big Idea

The argument in one line.

Garry Tan compressed an entire engineering org into 23 terminal slash commands, and the most important one runs before any code is written.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A developer or technical founder who already uses Claude Code and wants a structured workflow that prevents scope creep and skipped review phases.
An engineering manager curious about whether AI agent workflows can replace or replicate the accountability functions of a human team structure.
A solo builder who ships fast but regularly skips architecture review and wants a framework that forces those conversations before writing code.

SKIP IF…

You work in a large organization where real code review, QA, and release management involve compliance, regulated systems, or multi-team sign-offs — this is optimized for solo-to-small-team workflows.
You're looking for a tool-agnostic framework; gstack is tightly coupled to Claude Code and Anthropic's model stack, and most commands don't port cleanly elsewhere.

TL;DR

The full version, fast.

Garry Tan open-sourced gstack, a 23-slash-command framework that compresses a full engineering org chart into a terminal running inside Claude Code and nine other coding agents. The key architectural decision is that the first command is not a build instruction — it is /office-hours, which forces six scoping questions before any code is written, enforcing the process-beats-prompts discipline the repo is built around. Each subsequent command covers a distinct phase: planning, design review, code audit, QA, and automated ship-to-PR. The repo itself was built almost entirely by Claude, making it one of the most honest examples of AI-native development published this year. Critics note that Tan's platform amplifies everything he ships, but the workflow stands independently of the celebrity.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:13

01 · Cold open, credibility hook

Typing-hands B-roll. Garry Tan is YC CEO. GitHub card: 100K stars, 14.9K forks. Spoken undercut: it is not the code.

00:13 – 00:52

02 · 23 roles, not prompts

Animated org chart builds CEO to QA. Reframe: not prompts, not configs, roles. Portable skill layer across 10 agents.

00:52 – 01:08

03 · The command nobody writes first

Most agent setups: describe plus model writes code. GStack does something different. Sets up the /office-hours reveal.

01:08 – 01:52

04 · /office-hours and the 7-phase loop

Six forcing questions on scope, users, trade-offs, on the record. 7-phase wheel animates. /auto-plan chains the entire pipeline.

01:52 – 02:15

05 · Planning commands deep-dive

/plan-ceo-review, /plan-eng-review, /plan-design-review. Actual SKILL.md contents shown from GitHub.

02:15 – 02:51

06 · Building side commands

Terminal mockup: /review (staff audit, 12 fixes), /qa (browser-driven, 3 bugs), /ship (full release one command), /cso (OWASP + STRIDE).

02:51 – 03:21

07 · The 810x claim

Bar chart: 14 logical LOC/day in 2013 vs 11,400 in 2026. Caveat surfaced: logical not raw. 10-15 parallel sprints.

03:21 – 03:53

08 · Critics and caveats

@halo_check: YC amplifies. @opinionation: opinionation cuts both ways. Augment: study before installing.

03:53 – 04:39

09 · Meta-punchline

434 commits, 10 humans, 89 Claude co-authors across every Anthropic model. The repo teaching you to ship with AI was shipped by AI.

Atomic Insights

Lines worth screenshotting.

The first command in gstack is not /build — it's /office-hours, which forces you to answer six scoping questions before engineering starts.
Garry Tan compressed a 23-person engineering org into 23 slash commands that run inside a terminal.
The repo that teaches AI-native development was itself built almost entirely by AI — 89 of its coauthors are Claude model versions.
Process beats prompts: gstack's value is not in what you say to the AI but in the order you're forced to say things.
Tan claims 11,400 logical lines of code per day — 810x his 2013 rate of 14 — but uses logical LOC because AI inflates raw counts.
Running 10 to 15 sprints in parallel is only possible if handoffs between phases are enforced by structure, not personal discipline.
A single command, /auto-plan, chains four review roles in sequence — CEO, architect, designer, and engineer — with one keystroke.
The security review (/cso) runs OWASP and STRIDE threat modeling automatically — not as a checklist, as a slash command.
Critics point out that 100,000 stars on gstack partly reflects Garry Tan's platform, not purely the quality of the workflow.
Augment Code's review advice: study the repo before you install it — opinionated frameworks fight you when your workflow doesn't match the author's.
Karpathy admitted publicly he hasn't typed a line of code since December; gstack is what that admission looks like written as a system.
The /ship command syncs main, runs tests, audits coverage, and opens the PR — the entire release process in one terminal entry.

Takeaway

An Org Chart Compressed Into a Terminal

What it teaches

GStack replaces a 23-person engineering team with opinionated slash commands — the key insight is that roles and process, not raw prompts, are what make AI agents reliable.

01Cold open, credibility hook

A repo that reaches 100,000 stars in under two months without splitting the community is a signal worth paying attention to — the question is what it actually did differently.

0223 roles, not prompts

The value in an AI engineering setup is not the prompts or configs but the roles — each command models a specific organizational function with its own scope and expectations.
A portable skill layer that runs across ten different coding agents means the workflow is not locked to a single tool or model.

03The command nobody writes first

The most important command in any agent workflow is not the one that builds — it is the one that forces you to define scope, users, and trade-offs before a single line of code is written.

04/office-hours and the 7-phase loop

Requiring answers to six forcing questions on the record before engineering begins is what separates a disciplined agent workflow from a vibe-coding session.
A seven-phase loop — think, plan, build, review, test, ship, reflect — turns AI assistance into a repeatable process where the output of each phase becomes the input to the next.

05Planning commands deep-dive

Planning commands that challenge scope, lock architecture with edge cases, and rate design dimensions on a numeric scale replicate the pushback a real senior team would give.

06Building side commands

A single /ship command that syncs main, runs tests, audits coverage, and opens a PR demonstrates what it looks like to delegate an entire release workflow to an agent.
Security review — OWASP threat modeling and STRIDE analysis — can be a slash command rather than a separate engagement, making it cheap enough to run on every build.

07The 810x claim

An 810x productivity claim based on logical lines of code per day is worth scrutinizing: AI inflates raw LOC, and the metric only holds up if you define what you are measuring.

08Critics and caveats

A framework this opinionated fits well when your workflow matches the author's and creates friction when it does not — study a repo before installing it.
Platform amplification affects adoption numbers: the same repo from an unknown account would not have hit the same star count, which is a legitimate caveat to hold alongside the technical merits.

09Meta-punchline

The most honest proof of an AI-assisted workflow is a commit log: 89 AI co-authors across every model version is the repo proving its own thesis in public.

Glossary

Terms worth knowing.

gstack: An open-source GitHub repository containing 23 opinionated Claude Code slash commands that model a full software engineering organization, from CEO-level planning reviews to automated ship-to-PR workflows.
Slash command (Claude Code): A custom shorthand instruction prefixed with a forward slash that triggers a predefined workflow or role within Claude Code, allowing complex multi-step tasks to be initiated with a single terminal entry.
Claude Code: Anthropic's terminal-based AI coding assistant that operates within a project directory, reads and edits files, runs commands, and can be extended with custom slash commands and configurations.
Logical lines of code (LLOC): A measure of software complexity that counts meaningful executable statements rather than raw line count, filtering out blank lines, comments, and boilerplate that inflate simple metrics.
OWASP: The Open Web Application Security Project, an organization that publishes a widely-used list of the most critical web application security risks; the OWASP Top 10 is a standard checklist for security reviews.
STRIDE: A threat modeling framework that categorizes security risks into six types: Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, and Elevation of privilege.
Dogfooding: The practice of using your own product internally before releasing it publicly, treating your team as the first test user — derived from the phrase 'eating your own dog food.'

Resources

Things they pointed at.

00:10linkgarrytan/gstack ↗

01:39linkPulumi write-up on gstack

03:21linkAugment code review of gstack

Quotables