What gstack Actually Is
gstack is an open-source skill pack for Claude Code that turns a single Claude instance into a virtual engineering team. It ships as a ~/.claude/skills/gstack/ install that bundles 28 specialized slash commands, each activating a different "cognitive mode" inside Claude.
It was open-sourced by Garry Tan (CEO of Y Combinator) in March 2026. The repo hit 10,000 stars in 48 hours, then 66,000 stars in a few weeks — one of the fastest-growing developer tools on GitHub in 2026.
The thing to understand up front: gstack is Garry's actual setup. It's not a reference implementation or a "best practices guide." It's the opinionated tooling a specific senior operator uses every day. That's both why it's valuable and why it has detractors — you're inheriting Garry Tan's workflow opinions, whether or not they match yours.
The Roles gstack Creates
Each of the 28 slash commands pushes Claude into a different role. The headline ones:
- CEO (
/ceo) — rethinks the product from the user's perspective. Why does this feature exist? Is there a simpler version that achieves 80% of the outcome? - Designer (
/design) — catches generic AI output. Specifically designed to reject "looks like every other AI-built site" aesthetics. - Eng Manager (
/em) — locks architecture decisions before code is written. Asks hard questions about trade-offs, data model, failure modes. - Reviewer (
/review) — production-minded code review. Finds bugs that would ship but wouldn't survive first week of real traffic. - QA Lead (
/qa) — actually opens a browser (via Chrome DevTools Protocol) and tests the thing. Not simulated testing. - Security Officer (
/security) — runs OWASP Top 10 + STRIDE threat modeling against the code you just wrote. - Release Engineer (
/ship) — writes the PR description, handles the commit message, triggers the merge. - Doc Engineer (
/docs) — writes the docs alongside the code, not as an afterthought.
Plus ~20 more specialized commands for narrower tasks.
What Makes It Different From Just Using Claude Code
The short version: gstack forces context-switching discipline.
Without gstack, a typical Claude Code session has you asking one assistant to do many jobs simultaneously: "build this feature, but also make sure it's secure, and also think about the UX, and also write tests." Claude tries, and does a mediocre version of all of them.
With gstack, you explicitly sequence the jobs: /em to lock architecture, then build, then /security to audit, then /qa to test, then /ship to release. Each mode is more focused than "general Claude" and gets better output on its narrow task.
The cost is discipline. You have to actually invoke each mode, which means you have to actually think about which jobs apply before you ship.
The 10K-LOC/Week Claim
Tan has publicly said he uses gstack to ship ~10,000 lines of working code per week, solo. That number got people's attention — both as aspiration and as an argument against gstack (critics argue it's enabling "slop at scale").
My read from actually using it for a week: the 10K LOC number is plausible for someone who treats shipping as their primary job and has Garry's conviction to merge fast. For most engineers, the benefit isn't raw volume — it's that gstack turns a 3-hour "figure it out, code it, review it, ship it" cycle into a 45-minute sequence of focused steps. The volume gain is a side effect of reduced decision fatigue.
Why There's Hater Discourse
The criticism of gstack comes in three flavors, and all three have some merit:
-
"It's enabling AI slop at scale." Partially true. If you use gstack to ship 10K LOC/week of stuff nobody needs, you're shipping 10K LOC/week of slop. The tool doesn't check the product's value for you.
-
"It's Garry's workflow, not THE workflow." True. The CEO/EM/Designer hierarchy is a specific operator's mental model. A designer-led founder might want a different command palette. A research engineer doesn't need most of the modes at all.
-
"It over-abstracts what's already a simple tool." Partial. Claude Code is already a good general coder — adding 28 slash commands does create cognitive overhead. If you'd use 5 of the 28, pruning gstack to just those is likely correct.
None of these criticisms are disqualifying. They're reasons to use gstack intentionally, not reflexively.
Installing gstack (30 seconds)
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack
./setup
MIT licensed. Everything lives in .claude/. Nothing touches your PATH or runs in the background. Uninstall = rm -rf ~/.claude/skills/gstack.
Who Should Actually Install It
Good fit:
- Solo founders or 2-3 person teams who want large-organization discipline without headcount
- Engineers who've been burned by skipping review/QA steps and want a forcing function
- People shipping to real users (not just prototypes) where the cost of a bug is nonzero
Bad fit:
- Research/exploration work where you want Claude to iterate fast without ceremony
- Teams with an existing code review and QA workflow — gstack will fight your process
- People who learn better by building their own discipline rather than inheriting someone else's
How gstack Compares to Other Claude Code Plugins
gstack is methodology-first. You're inheriting Garry Tan's entire engineering philosophy.
Superpowers is narrower and more principled — its entire thesis is TDD + YAGNI. If you agree with those two principles, Superpowers is enforcing a cleaner philosophy than gstack.
claude-mem is orthogonal — it's memory infrastructure, not methodology. You can run it alongside gstack.
Anthropic's official code-review + security-guidance plugins are more modular — install just the review pieces without inheriting a full workflow.
If you want a complete opinionated workflow in one install: gstack. If you want composable pieces: the Anthropic plugins.
Can You Write Your Own Version?
Yes. gstack is 28 skill files + some glue. Looking at the source, each "role" is a markdown file with:
- A persona description
- A checklist of what that role checks/produces
- Output format rules
- When to hand off to the next role
If you don't like Garry's specific personas — the CEO/EM/Designer structure — you can write your own. The investment is ~2 hours per role. The Claude skills library has 2,300+ free skills you can model your own files after.
FAQ
Is gstack the same as Claude Code?
No. Claude Code is the tool. gstack is a skill pack you install into Claude Code.
Does gstack work with Cursor or other tools?
gstack is designed for Claude Code specifically — the skill file format and slash command registration are Claude Code conventions. Other tools have their own equivalents but gstack's files won't drop-in work.
Does Garry actually use this, or is it a marketing exercise?
By his public statements and the commit history, yes — he uses it daily. The repo is active. The personas match his public writing about how he thinks about product, engineering, and design.
Why does it have both 66K stars and loud critics?
Any highly opinionated tool attracts both. The stars come from people who want someone senior to tell them how to structure their workflow. The critics come from people who don't want that particular senior's opinions imposed on their work. Both are reasonable takes.
Should I install gstack OR Superpowers?
If you already have a clear engineering philosophy and want tighter discipline around it — Superpowers. If you want to inherit Garry Tan's full operational style — gstack. They overlap enough that installing both is usually redundant.
Does gstack make the Claude Certified Architect exam easier?
Indirectly. Domain 3 (Claude Code Configuration & Workflows) is 20% of the CCA exam. Having installed and used a real plugin — any of them — teaches you skill file format, slash command design, and hook mechanics in a way reading docs doesn't.
The Short Version
gstack is the first Claude Code plugin that feels like it was shipped by someone who uses it every day at scale, not someone demoing a feature. That's what makes it worth installing — and what makes it divisive. You're not getting "best practices." You're getting Garry Tan's practices. Whether that's a win depends on how much you want to trade autonomy for speed.