Launching today
GStack
Use Garry Tan's exact Claude Code setup
350 followers
Use Garry Tan's exact Claude Code setup
350 followers
gstack turns Claude Code from one generic assistant into a team of specialists you can summon on demand. Six opinionated workflow skills for Claude Code. Plan review, code review, one-command shipping, browser automation, and engineering retrospectives — all as slash commands.
The slash command approach resonates with how we've been running Claude Code autonomously. We built 100+ SKILL.md files over 108 hours of unattended operation and the key insight was exactly this: specialized roles outperform a single generic agent. The biggest win was separating "plan review" from "code review" — when Claude tries to do both in one pass, it either rubber-stamps or gets lost in details. Question for the team: does GStack handle context window pressure when running multiple skills in sequence? That's been our main scaling challenge.
@yurukusa The context window pressure with sequential skills is real — we hit the same wall. The other layer we found: even with perfect skills, the agent still doesn't know why the codebase is structured the way it is. That lives in tribal knowledge, not in any SKILL.MD. Curious whether your 100+ skills encode any of that historical reasoning or if it's purely procedural?
I've been using Claude Code with MCP for about a year now and keep iterating on my setup. Curious what's in Garry's stack — is it mostly about the MCP server configuration or also specific prompt engineering patterns? The biggest unlock for me was getting file system + browser MCP tools working together. Does GStack cover the agentic workflow side or is it primarily IDE-focused?
The sub-200ms persistent Chromium daemon is the real unlock here - eliminates the cold start penalty that makes browser automation painful in agentic workflows. The separation between plan review and code review as distinct slash commands is smart - keeps Claude focused on one cognitive task at a time. Curious if there's state persistence between commands for iterative debugging sessions.
MacQuit
This is exactly the direction Claude Code needs to evolve. I've been building native macOS apps (SwiftUI + AppKit) entirely with Claude Code and the biggest lesson matches what GStack is doing — you need structured roles, not a single "do everything" agent.
My setup uses a similar pattern: separate skills for planning, implementation, and code review, plus a CLAUDE.md that enforces TDD and incremental commits. The game changer was adding hooks (PostToolUse, Stop) for automated checks — so the agent can't ship code that doesn't compile or pass linting.
One thing I'd love to see in GStack: a "native app" skill that understands platform-specific constraints. Web devs can iterate fast with hot reload, but when you're building for macOS/iOS, the feedback loop is longer and the agent needs to be more cautious about breaking changes.
The headless browser piece with bun runtime is clever — that's been a gap in most setups. Great launch!
Congrats on the launch! GStack is exactly what the Claude Code ecosystem needed — role-based specialization instead of one generic mode.
One thing worth adding to this setup: a seatbelt layer.
We ran GStack-style autonomous sessions and hit Case #001: a Claude Code agent looped for 70 minutes, repeatedly injecting a staging URL into a production config file. Every log showed exit code 0. All green. The deviation was invisible because nothing had recorded what the agent intended to do before it acted.
Built K9 Audit to fix this — a deterministic, non-LLM causal auditing layer that drops directly into .claude/settings.json (zero code changes, fully compatible with GStack). It records a cryptographically hashed 5-tuple of declared intent vs actual outcome. When something goes wrong, k9log trace --last gives root cause in under a second.
GStack for velocity. K9 Audit as the seatbelt.
Repo: https://github.com/liuhaotian2024-prog/K9Audit
Why is this on Product Hunt?
This is a bunch of prompts. You have got to be kidding.
Y Combinator
@sherveen Actually I use a bun runtime and it does a lot to make headless browser useful. This one for instance will import your login cookies securely from ALL the browsers on your Mac.
@garrytan Garry, let's be clear and honest: if you weren't the CEO of YC, this wouldn't be on PH.
Birdhouse
I resonate with this so hard. Agents (like humans) can't perform as well when they wear too many hats. The future is coordinating teams of specialized agents like this with just the right amount of human in the loop.
It's a freeing feeling when your workflows reach this level of parallelization. The productivity gains are insane.
Can't wait to try these out in Birdhouse. Thanks for sharing!