gstack 是什么:YC 总裁 Garry Tan 把 Claude Code 拆成 23 个 slash 命令,108k stars 的 skill 套装实测
一条推特营销贴把 YC 总裁 Garry Tan 的 garrytan/gstack 仓库推上风口。"23 ready skills"、"3 commands. 1 markdown file."、"108k stars"、"an 18-year-old won a hackathon"。营销腔掩盖了仓库本身的信息密度。
把推特原贴和 README 对照之后,可核实事实和营销话术可以分开。
一条推特的原话
Ronin (@deronin_) 的核心话术:
- 3 commands. 1 markdown file. 23 skills
- 108k stars. made by [president of YC]
- /office-hours — YC-style review before you write a line of code
- /plan-eng-review — full architecture from one product decision
- /design-shotgun — 4-6 design variants in parallel, opens in your browser
- /qa — opens a real browser and clicks through your flows
- GBrain — persistent memory between sessions
- Solo 跑营销 agency、zero employees、mornings 现在一小时搞定
- 营销话术:"80% of [Claude Code] potential on the table"、"cheat code for planning"
仓库核验
到抓取时刻,api.github.com/repos/garrytan/gstack 返回:
- Stars: 108,048
- Forks: 16,071
- 描述: "Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA"
- License: MIT
- 作者背景(README 自述): Garry Tan is President & CEO of Y Combinator. Palantir 早期 eng/PM/designer;cofounded Posterous(卖给 Twitter);built Bookface(YC 内部社交网络)
Garry Tan 自报产出数据(README 引用,方法学链接 docs/ON_THE_LOC_CONTROVERSY.md):
- 2026 vs 2013 logical code change rate: ~810×(11,417 vs 14 logical lines/day)
- 2026 YTD 已经是 2013 全年的 240×
- 覆盖 40 个 public + private
garrytan/*仓库,排除一个 demo repo - README 主动承认 raw LOC 在 AI 时代被通胀——所以用了"logical"归一化口径
README 开头引用 Karpathy 在 No Priors 播客 2026-03 那段:
"I don't think I've typed like a line of code probably since December, basically, which is an extremely large change."
gstack 实际是什么
Garry 自己一句话定义:
"It turns Claude Code into a virtual engineering team — a CEO who rethinks the product, an eng manager who locks architecture, a designer who catches AI slop, a reviewer who finds production bugs, a QA lead who opens a real browser, a security officer who runs OWASP + STRIDE audits, and a release engineer who ships the PR. Twenty-three specialists and eight power tools, all slash commands, all Markdown, all free, MIT license."
本质:把"自己验证过的工作流"打包成 23+ 个 Claude Code slash command + 8 个 power tool,加上 memory(GBrain)和 OpenClaw 集成,全部 Markdown、全部 MIT 协议、全部免费。
6 步上手(README 推荐路径)
- Install gstack(30 秒)
- 跑
/office-hours描述要建什么 - 跑
/plan-ceo-review过任一 feature idea - 跑
/review过任一有改动的 branch - 跑
/qa跑 staging URL - 停在这。你会知道这工具对你合不合适
安装流(30 秒)
requirements:Claude Code、Git、Bun v1.0+、Node.js(仅 Windows)。
一行 prompt 让 Claude 替你装:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack
./setup
setup 后 README 给的完整 skill 列表(节选):/office-hours、/plan-ceo-review、/plan-eng-review、/plan-design-review、/design-consultation、/design-shotgun、/design-html、/review、/ship、/land-and-deploy、/canary、/benchmark、/browse、/connect-chrome、/qa、/qa-only、/design-review、/setup-browser-cookies、/setup-deploy、/setup-gbrain、/retro、/investigate、/document-release、/document-generate、/codex、/cso、/autoplan、/plan-devex-review、/devex-review、/careful、/freeze、/guard、/unfreeze、/gstack-upgrade、/learn。
实际不止 23——setup 列表已经数到 30+ 个,营销帖的"23"是低估。
团队模式
在 repo 内跑 ./setup --team && gstack-team-init required,自动提交 .claude/ + CLAUDE.md。没 vendor 文件进 repo、没版本漂移、没手动升级——每次 Claude Code session 启动时跑一次 auto-update check(节流到 1 次/小时,网络失败静默)。
required 可换 optional——required 会 block 不兼容的 teammate,optional 只 nudge。
OpenClaw 集成
OpenClaw 通过 ACP spawn Claude Code session,所以 Claude Code 装了 gstack 之后,所有 gstack skill 在 OpenClaw agent 里直接能用。
README 还提供 4 个原生 OpenClaw skill(直接走 ClawHub,不开 Claude Code session):
clawhub install gstack-openclaw-office-hours gstack-openclaw-ceo-review gstack-openclaw-investigate gstack-openclaw-retro
| Skill | 用途 |
|---|---|
gstack-openclaw-office-hours | 6 个 forcing question 逼你把产品想清 |
gstack-openclaw-ceo-review | 4 种 scope 模式做战略挑战 |
gstack-openclaw-investigate | 根因 debugging 方法论 |
gstack-openclaw-retro | 每周工程回顾 |
这些是对话型 skill——OpenClaw agent 直接在 chat 里跑。
不止 Claude Code:8 个 host
./setup auto-detect 已装的 agent。也可用 ./setup --host <name> 显式指定:
| Agent | Flag | 装到哪 |
|---|---|---|
| OpenAI Codex CLI | --host codex | ~/.codex/skills/gstack-*/ |
| OpenCode | --host opencode | ~/.config/opencode/skills/gstack-*/ |
| Cursor | --host cursor | ~/.cursor/skills/gstack-*/ |
| Factory Droid | --host factory | ~/.factory/skills/gstack-*/ |
| Slate | --host slate | ~/.slate/skills/gstack-*/ |
| Kiro | --host kiro | ~/.kiro/skills/gstack-*/ |
| Hermes | --host hermes | ~/.hermes/skills/gstack-*/ |
| GBrain(mod) | --host gbrain | ~/.gbrain/skills/gstack-*/ |
这是一个 host-agnostic 框架——skill 描述用 markdown,不绑 Claude Code 内部 API。
一段 session 长这样(README 摘录)
You: I want to build a daily briefing app for my calendar.
You: /office-hours
Claude: [asks about the pain — specific examples, not hypotheticals]
You: Multiple Google calendars, events with stale info, wrong locations.
Prep takes forever and the results aren't good enough...
Claude: I'm going to push back on the framing. You said "daily briefing
app." But what you actually described is a personal chief of
staff AI.
[extracts 5 capabilities you didn't realize you were describing]
[challenges 4 premises — you agree, disagree, or adjust]
[generates 3 implementation approaches with effort estimates]
RECOMMENDATION: Ship the narrowest wedge tomorrow, learn from
real usage. The full vision is a 3-month project — start with
the daily briefing that actually works.
注意 /office-hours 做的事:先 push back framing,再问 6 个 forcing question,再给 3 个 implementation approach。这不是 single-shot prompt,是带方法论的对话。
营销 vs 事实
Ronin 推特里的可核实部分:
- 23+ skills ✅(实际 30+)
- 3 commands 装好 ✅(clone + setup + CLAUDE.md)
- 108k stars ✅
- President of YC ✅
- /office-hours、/design-shotgun、/qa、GBrain、document-release 都存在 ✅
- 8 个 host 支持 ✅
- OpenClaw ACP 集成 ✅
营销腔需要折扣:
- "cheat code for planning" — 主观偏好
- "80% of [Claude Code] potential on the table" — 无 benchmark
- "an 18-year-old won a hackathon" — 轶事,README 不可验证
- "ships in hours what used to take days" — 个人声称
- "zero employees" 跑营销 agency — Ronin 自己,不是 Garry
6 条冷静判断
- skill 套装的本质:把"经过自己验证的工作流"打包成 slash command 分发。同领域持久化 agent 框架也都有 skill 系统,gstack 的差异是把这套做法外部化、公开化、跨 host 化
- host-agnostic 设计的价值:8 个 host 共享同一套 skill 定义,等于把方法论做成 portable asset。不会被某个 AI 工具的兴衰绑定
- QA in real browser 是最后一公里:agent 能写代码不算稀奇,能跑 browser 自动化 QA 找 bug 才是。从
/connect-chrome到/qa这一段是 gstack 真正难写的部分 - persistent memory 跟其它项目殊途同归:GBrain 解决"agent 跨 session 忘事",和当前主流 agent 框架的 memory 系统解决的是同一类问题——工程化路径不同,目的相同
- Karpathy 引用 + Garry Tan 自报数据 + 108k stars:这三个互相佐证,让"810× productivity"这种夸张数据不全是营销话术,但需要读
ON_THE_LOC_CONTROVERSY.md自己核验方法学 - "marketing agency solo 跑"是 Ronin 的故事不是 Garry 的:Garry 自己公开的数据是 YC 总裁 + 副业 ship 40+ features;Ronin 套到自己 marketing agency 上是另一回事。读者别把这两件事当同一件事
该不该装
/office-hours 走 6 个 forcing question + push back framing——如果你发现自己经常 ship 完才发现"想错了"再返工,这套方法论的 ROI 可能是真的。
/qa 跑真实 browser——如果你测试全靠手动或者 unit test,从 staging URL 自动化 QA 这段是肉眼可见的省时间。
GBrain 跨 session memory——如果你试过 Claude Code 重启后让 agent 接着昨天的 context,它记得约不了一件事就知道痛点。
最不容易判断的是 23 个 skill 一起装的边际效用。README 的建议是先跑 quick start 那 6 步——/office-hours → /plan-ceo-review → /review → /qa——觉得合适再继续。这条建议比营销帖所有话术都更可信。