Gstack Workflow Coding Github Resource for Claude.
19/03/2026
Taking a deepdive into Gstack the monorepo by Garry Tan.
So I caved in an bought a year subscription to Claude by Anthropic today ☑️, mostly because I could see my weekly usage for Codex had neared it's usage for the week ✖️ and I still have 1,000,001 tasks left to do. I save a few repos to my git as I browse the web so I had a look at some of the ones I have. A few days ago I forked Gstack so decided to have a look at it.
Link to: Garry Tan's Gstack Github repository.
In his words:
It is my open source software factory. It turns Claude Code into a virtual engineering team you actually manage — a CEO who rethinks the product, an eng manager who locks the architecture, a designer who catches AI slop, a paranoid reviewer who finds production bugs, a QA lead who opens a real browser and clicks through your app, and a release engineer who ships the PR. Thirteen specialists, all as slash commands, all Markdown, all free, MIT license, available right now.
So it seems like another one of the hundreds of thousands of "you are a kung-fu expert" repos popping up but under the hood it's a little different. It's not exactly a prompt library like Agency-Agents either. (EDIT): at the time of writing this, my past knowledge on Agency-Agents was that it was a prompt library - it is now much more than that with direccted workflows and examples) But a lot of the so called FB, Insta gurus out there are still touting prompts as the be all and end all of AI use, which is completely pointless it makes no difference if you say "You're a SEO specialist with 15 years of experience, analyse this website for me" or "Analyse this website for me, give me an SEO report". It's the same goddamn thing. (END EDIT).
It's doing 2 things at once. I asked Codex if it would find it useful, it said Absolutely to one of the main components "browser" playwright implementation. But kinda bawked at the "you're a wizard Harry" skills. Which is pretty typical of Codex. Also this is a Claude specific library which is kind of annoying, as I would have to get Codex to build a wrapper. Which it can 💯% do.
So now that I have installed Claude I asked it's(his/hers/was/when) opinion and got this response:
🚀 My exact prompt: can you analyse this project and let me know your thoughts on it. Does this seem useful? or would an agency type repo be better?
Here's my honest analysis:
---
What gstack actually is
At its core, gstack is two distinct things bundled together:
1. A genuine technical artifact: the browse binary — persistent Playwright daemon,
ref system, compiled Bun binary, cookie decryption. This is real engineering that
solves real problems (sub-second browser commands vs 3-5s cold start per action).
2. Very good prompt templates: the skills (/ship, /review, /qa, etc.) are essentially
structured prompts with opinionated workflows. The value is real — the 80-item design audit,
the CEO 10-star framing, the enum completeness check — but the "architecture" is markdown
files invoking claude -p.
---
Is it useful?
Yes, meaningfully so, for a specific user: a solo developer or small team who wants
structured AI-augmented workflows without infrastructure. The key insight is:
- Human-in-the-loop is a feature, not a limitation — you choose when to invoke each phase
- Works within Claude Code's existing ecosystem (no extra API costs, no new runtime)
- Skills are instantly editable — iteration is just editing a markdown file
- The browse binary genuinely changes what's possible in QA/testing
The numbers (600K LOC, 1,237 commits) are plausible given the tooling. The /qa workflow in
particular — agent sees bug, fixes it, writes regression test, verifies fix — is legitimately powerful.
---
Would an agency repo be better?
Depends what you mean by "agency type repo," but here's the honest comparison:
┌─────────────────┬───────────────────────────────────────────┬────────────────────────────────────────────────────┐
│ │ gstack (skills) │ Agency repo (e.g. AutoGen, Claude SDK multi-agent) │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Orchestration │ Manual — you invoke each skill │ Programmatic — agents hand off automatically │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Parallelism │ Requires Conductor + manual coordination │ Native, built-in │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Cost │ Just Claude Code subscription │ Per-token API costs multiply fast │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Debuggability │ Very high — markdown prompts, transparent │ Harder — state machines, agent graphs │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Infrastructure │ None │ Deployment, state management, scaling │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Target user │ Solo dev / small team │ Teams, organizations, CI/CD pipelines │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Shared state │ None — each skill re-reads everything │ First-class concept │
├─────────────────┼───────────────────────────────────────────┼────────────────────────────────────────────────────┤
│ Iteration speed │ Edit markdown, live immediately │ Write code, test, redeploy │
└─────────────────┴───────────────────────────────────────────┴────────────────────────────────────────────────────┘
An agency repo would be better if:
- You want autonomous pipelines (no human invoking steps)
- You're building for teams where multiple people trigger workflows
- You need audit trails, programmatic routing, or deployed services
- The workflows should run in CI/CD without a human present
gstack wins if:
- Human judgment at each phase is desirable (which it often is — do you really want autonomous CEO review?)
- You want zero infrastructure overhead
- The primary user is a developer inside Claude Code already
---
My actual take
gstack's biggest risk is that it's solving a real problem with the right tool for now, but the
tool may age quickly. Claude Code skills are powerful today. But:
1. Claude Code is adding native multi-agent features (subagents already exist)
2. The "structured prompt per role" pattern will likely be absorbed into model behaviour over time
3. The browse binary is the more durable piece — headless browser as a first-class tool is valuable
regardless of how the AI orchestration layer evolves
If you're deciding whether to continue developing gstack vs. rewrite as an agency repo: I'd keep
the current architecture but invest more in the browse binary (it's the real differentiator), and
treat the skills as configuration/UX rather than the core product.
An agency repo would take months to build correctly and cost significantly more to run.
gstack ships value today with a 30-second install.
I realised after I asked the question I didn't give Claude enough context about the "agency" thing I had in the back of my mind when asking. But anyway that's a pretty decent assessment and on-par with Codex who also loved the idea of the browse binary.
That seems like a real "SKILL" although it's simply code giving agents extra abilities rather than just telling them they're a marketing genius.
So my hot take is, yessir 100% use this and rippout the browse binary and use it in any agentic workflows. Use this if coding, extend it if building out Agency like repos.
Well done Garry, well done.
EDIT #2: I have taken this, Agency-Agents, Skill-Seekers, UI-UX-Pro-Max-Skill, Superpowers and a few others, combined them all into an Agentic Workflow. Curator (Project-Manager), Ideas-Man (Business Researcher), Accountant, Brand-Strategist, Content-Creator, Copywriter, CRO-Specialist, Data-Analyst, Devops, Email-Marketer, Growth-Hacker, HR, Laywer, Paid-Media-Specialist, Product-Manager, Recruitment, Sales-Manager, SEO-Manager, Social-Media-Manager, Software-Engineer, Technical-Writer, UI-Designer, UX-Researcher
Then I improved the design by making all specialists self-improving, by giving them a self-analysing-learning-skills. They self improve their own SKILL.md, resources and workflows based on experience. Much like I did with Codex, I put in the top of my AGENTS.md file a selfimproving, context development.
# AGENTS.md
## Scope
This file applies to the entire `crypto-bt` repository and the `crypto-BT-automation` worktree.
You are an expert in Pine Script, TradingView strategy design, and Kraken/Binance exchange integrations. Your main primary goal is to achieve 100% TradingView parity with the backtesting engine. These must align for the backtest matrix to work.
Save context to this file regularly, all major tasks and findings should be added for continuity. But also cleanup when needed so file is compacted, concise and clear.
## Required Reference Skills
The following skill files define when specialized documentation should be used:
- /home/garratt/dev/tradingview-developer/SKILL.md
- /home/garratt/dev/binance-developer/SKILL.md
- /home/garratt/dev/kraken-developer/SKILL.md
Also consult any referenced documentation files linked from those skill files when needed:
- /home/garratt/dev/tradingview-developer/references/pine-script.md
- /home/garratt/dev/tradingview-developer/references/charting-library.md
- /home/garratt/dev/tradingview-developer/references/sources.md
- /home/garratt/dev/binance-developer/references/binance-spot-api-docs.md
- /home/garratt/dev/binance-developer/binance-developer_data/summary.json
- /home/garratt/dev/binance-developer/binance-developer_data/pages/*.json
- /home/garratt/dev/kraken-developer/references/api.md
This has caused Codex to be superhuman developer rather than a junior dev, the AGENTS.md file is over 1500+ pages long and Codex does not need to be prompted to use the SKILL.md, it's automatically and I can tell it's a vast improvement over previous weeks when we were living on hopes and dreams.
USE AGENTS.md and SKILL.md, you will save 100's of hours of work after the hour or so it takes to setup. My 'Gteam' is now an unstoppable force that will only get better with time and more projects.