WEBVTT 00:00:00.080 --> 00:00:03.360 Gary Tan is the CEO of Y Combinator. 00:00:03.440 --> 00:00:10.240 He just open sourced his AI engineering team, all 23 of them. The part that makes it work isn't the code. 00:00:10.400 --> 00:00:13.280 The repo is Gary Tan slash g stack. 00:00:13.775 --> 00:00:14.735 100,000 00:00:14.735 --> 00:00:22.815 stars in under two months. And here's the unusual part. Repos this opinionated normally split the room. This one isn't splitting. 00:00:23.055 --> 00:00:29.660 So what did Tan ship that everyone's agreeing on? G Stack is 23 opinionated slash commands. 00:00:29.980 --> 00:00:32.140 Not prompts, not configs, 00:00:32.300 --> 00:00:33.020 roles. 00:00:33.260 --> 00:00:34.540 There's a CEO, 00:00:34.540 --> 00:00:37.900 a designer, an engineering manager, a release manager, 00:00:38.060 --> 00:00:39.980 a doc engineer, and QA. 00:00:40.485 --> 00:00:46.245 Each one is a command you type. They run inside Claude code plus nine other coding agents. 00:00:46.565 --> 00:00:52.165 The skill layer is portable. The pitch is simple. An org chart compressed into a terminal. 00:00:52.485 --> 00:01:01.600 And the most important command in it is the one nobody writes first. Most agent setups start the same way. You describe what you want. The model writes code. 00:01:01.920 --> 00:01:03.760 GStack does something different. 00:01:04.080 --> 00:01:08.320 The first command isn't slash build. It's slash office hours. 00:01:08.720 --> 00:01:26.050 Before any line of code, an AI playing the CEO asks you six forcing questions about scope, about users, about trade offs. You have to answer them on the record. Only then does engineering get to start. Then the loop kicks in. Think, plan, build, 00:01:26.130 --> 00:01:26.930 review, 00:01:27.090 --> 00:01:27.810 test, 00:01:27.970 --> 00:01:28.690 ship, 00:01:28.850 --> 00:01:29.650 reflect. 00:01:30.050 --> 00:01:32.050 Each phase has its own command. 00:01:32.450 --> 00:01:37.010 The output of one is the input to the next, so handoffs don't get skipped. 00:01:37.490 --> 00:01:39.650 Palumi's write up summed it up in three words. 00:01:40.305 --> 00:01:42.145 Process beats prompts. 00:01:42.305 --> 00:01:47.265 And one command, slash auto plan, chains the entire review pipeline together. 00:01:47.665 --> 00:01:53.025 Type once. Get four roles reviewing in sequence. Let's open them. Look at the planning commands, 00:01:53.430 --> 00:02:01.510 Slash plan CEO review challenges your scope with four decision modes. What to cut, what to defer, what to commit to. 00:02:01.910 --> 00:02:05.030 Slash plan and review locks the architecture 00:02:05.190 --> 00:02:08.310 with data flow diagrams and edge cases listed out. 00:02:08.885 --> 00:02:15.845 And slash plan design review actually rates dimensions of your design on a zero to 10 scale. 00:02:16.165 --> 00:02:28.660 It's the kind of pushback you'd normally only get from a real team. That covers the planning side. The building side has its own commands. Slash review runs a staff engineer code audit and auto fixes the obvious stuff. 00:02:28.980 --> 00:02:32.500 Slash QA drives a real browser, finds bugs, 00:02:32.660 --> 00:02:34.340 writes regression tests, 00:02:34.660 --> 00:02:36.340 slash ship syncs main, 00:02:36.500 --> 00:02:37.620 runs tests, 00:02:37.860 --> 00:02:38.740 audits coverage, 00:02:39.115 --> 00:02:39.995 pushes, 00:02:40.075 --> 00:02:41.435 opens the PR. 00:02:41.755 --> 00:02:42.715 One command, 00:02:42.875 --> 00:02:43.835 full release. 00:02:43.915 --> 00:02:48.475 There's even a slash CSO that does OWASP and stride threat modeling, 00:02:48.715 --> 00:02:51.355 security review as a slash command. 00:02:51.675 --> 00:02:54.420 So does any of this actually make him faster? 00:02:54.660 --> 00:02:56.020 Tan says yes. 00:02:56.340 --> 00:02:57.860 11,400 00:02:57.860 --> 00:03:02.980 logical lines of code a day versus 14 back in 2013. 00:03:03.380 --> 00:03:04.340 810 00:03:04.340 --> 00:03:04.660 times. 00:03:05.205 --> 00:03:08.565 He also says he runs 10 to 15 sprints in parallel. 00:03:08.885 --> 00:03:12.085 The caveat, he uses logical LOC, 00:03:12.085 --> 00:03:12.965 not raw. 00:03:13.285 --> 00:03:16.005 Because in his words, AI inflates raw. 00:03:16.165 --> 00:03:21.610 Whether you buy that math is your call, the repo doesn't depend on it. Critics aren't quiet. 00:03:21.930 --> 00:03:26.170 Would the same repo from an unknown developer hit a 100,000 stars? 00:03:26.490 --> 00:03:27.530 Probably not. 00:03:27.930 --> 00:03:31.530 YC's platform amplifies everything tan ships, 00:03:31.930 --> 00:03:33.690 and opinionation cuts both ways. 00:03:34.385 --> 00:03:36.785 If your workflow matches TANs, 00:03:36.945 --> 00:03:38.705 g stack is a gift. 00:03:38.945 --> 00:03:41.345 If it doesn't, you'll fight the framework. 00:03:41.985 --> 00:03:44.225 The augment code review put it well. 00:03:44.705 --> 00:03:53.360 Study the repo before you install it. Still, Gstack is the operational version of a vibe a lot of senior engineers have been quietly admitting. 00:03:53.600 --> 00:03:55.920 Andre Carpathi said it out loud. 00:03:56.560 --> 00:03:57.200 Quote, 00:03:57.520 --> 00:04:01.280 I don't think I've typed like a line of code probably since December, 00:04:01.760 --> 00:04:02.160 unquote. 00:04:02.735 --> 00:04:07.935 Gstack is what that sentence looks like when you write it down as code. And the meta punchline. 00:04:08.175 --> 00:04:10.575 Open the Gstack commit log. 00:04:11.135 --> 00:04:12.415 434 00:04:12.415 --> 00:04:13.215 commits, 00:04:13.615 --> 00:04:15.215 roughly 10 human contributors, 00:04:15.740 --> 00:04:18.220 and 89 distinct coauthors, 00:04:18.300 --> 00:04:20.140 almost all of them clawed 00:04:20.140 --> 00:04:27.820 across basically every model version Anthropic released this year. The repo that teaches you how to ship with AI was itself shipped by AI. 00:04:28.060 --> 00:04:32.808 Most honest dogfooding you'll see this year, fork it, steal three ideas, 00:04:32.888 --> 00:04:38.248 or laugh at the LOC math. Just don't ignore. An org chart is now a slash command.