Garry Tan pushes skillifying workflows

AI coding is starting to split into two camps. One camp still treats the model like a very smart intern you brief from scratch every time. The other camp is trying to turn recurring work into reusable machinery. Garry Tan is very clearly in the second camp now — and the interesting part is that he is not just talking about it. He has open-sourced the actual stack he uses, with named skills, memory, and orchestration layers that make the workflow persistent instead of improvised. ### What does “skillifying” actually mean? Basically, it means you stop writing the same long prompt over and over. Instead, you package that prompt, its rules, and its expected output into a reusable skill. So instead of saying “please review this code like a paranoid staff engineer,” you invoke a review skill that always does that job the same way. That sounds small, but it changes AI work from chat into process. (github.com) ### Why is that better than ad-hoc prompting? Because ad-hoc prompting does not compound. A great prompt in a chat window disappears into scrollback. A skill can be versioned, reused, shared, tested, and improved. Teams can decide which behaviors they trust and keep those behaviors stable across projects. That is the real shift here — from “ask the model nicely” to “install a workflow primitive.” ### What did Garry Tan actually ship? (github.com) The clearest artifact is gstack, his public repo for AI-assisted software work. The repo describes 23 opinionated tools covering roles like CEO, designer, engineering manager, release manager, doc engineer, and QA. The structure itself tells the story: he is breaking software work into repeatable cognitive modes, then giving each mode a command surface. As of May 11, 2026, the repo shows about 93.4k stars and 13.8k forks. ### Where does GBrain fit in? GBrain is the memory layer. Tan’s repo describes it as an agent brain, and the project has its own skills, templates, evals, and sync flows. In plain English, gstack is the set of workflows, while gbrain is the place those workflows can store context, preferences, and retrieved knowledge over time. That pairing is the important idea — skills without memory are brittle, and memory without skills is just a pile of notes. (github.com) ### Why are people paying attention now? Because this is no longer a vague “agents will do stuff” pitch. It is attached to concrete open-source projects with visible adoption and rapid iteration. gbrain shows about 14.4k GitHub stars today, and both repos are updating constantly, with commits landing within hours. That makes the whole thing feel less like a keynote concept and more like a live operating system for AI-native work. (github.com) ### Is this just for coding? No — and that is where the YC angle matters. Tan has also been pushing a broader argument that small teams can beat much larger ones by automating internal functions across engineering, ops, and support. So “skillifying” is really a company design idea, not just a Claude Code trick. The bet is that startups should encode the way they work into reusable automations as early as possible. (github.com) ### What’s the catch? The catch is governance. Once workflows become durable, they also become policy. Someone has to maintain them, decide when they are wrong, and keep them from calcifying into bad defaults. A skill is like a macro with opinions — incredibly useful, but only if the team keeps auditing the opinions inside it. ### So what’s the real takeaway? Tan’s push is not “use better prompts.” It is “treat repeated AI work like software.” When a workflow gets used twice, maybe write a prompt. (ycombinator.com) When it gets used 20 times, turn it into a skill. When the whole company depends on it, give it memory, tests, and ownership. That is the deeper shift hiding underneath the demos. (github.com)