How to Build Software with GStack: AI Agent Workflow Tutorial

Name: Inside Garry Tan's Claude Code Setup
Uploaded: 2026-04-23T14:30:49+00:00
Duration: 21 min 49 s
Channel: Y Combinator
Description: Summary and key takeaways on Inside Garry Tan's Claude Code Setup: Summary & Key Takeaways, covering The Agent Era of Software Development Building software

Y Combinator

Apr 23, 2026

•

21 min video

•

2 min read

YouTube video ID: wkv2ifxPpF8

Source: YouTube video by Y Combinator — Watch original video

PDF

Building software with agents requires the same structure as human teams: defined roles, clear processes, and systematic reviews. The “thin harness, fat skills” approach keeps the scaffolding minimal while letting the model’s intelligence do the heavy lifting. As long as the models are set upright, they are already smart enough to produce extraordinary code, but without proper scaffolding the output tends to be plausible yet broken.

The GStack Framework

GStack is an open‑source repository that turns Claude Code into an AI engineering team. It supplies specialized skills that mimic human workflows. The Office Hours skill distills the Y Combinator partner methodology, forcing users to answer six key questions that reframe the product before any code is written. An adversarial review process lets the model critique its own design document, flagging missing components such as two‑factor authentication, privacy safeguards, or failure handling, and then attempts to auto‑fix them. The Design Shotgun tool generates multiple UI/UX concepts using OpenAI Codex image generation, giving the user a visual menu to choose from.

Workflow Demonstration: Building a Tax App

The tutorial starts with an Office Hours session that asks forcing questions to validate the tax‑app idea and its business model. After the idea passes this filter, the model runs a multi‑step adversarial review, raising the design‑doc score from 6/10 to 8/10 by identifying gaps and proposing fixes. Next, Design Shotgun produces several UI mockups; the user selects the preferred version, and the model proceeds to implement the chosen design. Throughout, the speaker emphasizes that “the way to get agents to do real work is the same way humans have always done it: as a team with roles, with process, with review.”

Scaling and Automation

GStack wraps Playwright and Chromium in a CLI tool called GStack Browser, enabling agents to log into websites, navigate pages, download PDFs, and run regression tests in a visible browser. The Ship tool prepares pull requests for merging into the main branch, automating the final integration step. With these utilities, the system supports “Level 7” software‑factory operations: multiple agent sessions run in parallel on different features or projects, allowing the processing of 10–50 pull requests per day.

Managing Parallel Development

The speaker manages 10–15 concurrent Claude Code sessions by creating a new work tree for each bug report or feature idea. Instead of a traditional to‑do list, the user simply initiates the standard GStack process—Office Hours, CEO Review, adversarial review, Design Shotgun, and Ship—for each item. This parallel PR workflow eliminates bottlenecks and keeps development fluid, turning the cloud into “just someone else’s computer” that can be harnessed at scale.

Takeaways

Treat AI agents like a human team by assigning roles, defining processes, and adding review steps to achieve reliable software output.
GStack provides a thin harness with specialized skills such as Office Hours, adversarial review, and Design Shotgun, enabling Claude Code to function as an engineering team.
The tax‑app demo shows the sequence of Office Hours validation, adversarial design‑doc improvement, and UI generation with Design Shotgun before code is written.
Scaling relies on the GStack browser CLI and the Ship tool, which let agents automate browsers and prepare pull requests, supporting dozens of parallel sessions and high‑volume PR handling.
Managing 10–15 concurrent Claude Code sessions is done by spawning separate work trees for each bug or feature, removing traditional to‑do lists and streamlining development.

Frequently Asked Questions

What is the 'Office Hours' skill in GStack and how does it shape product ideas?

Office Hours is a distilled version of the Y Combinator partner process that forces users to answer six critical questions, reframing the product idea before any code is generated. This early validation helps ensure the concept is viable and aligns with business goals.

How does GStack’s adversarial review improve a design document?

Adversarial review lets the model critique its own design doc, automatically spotting missing elements like two‑factor authentication, privacy safeguards, or failure handling. It then attempts to fix these gaps, raising the document’s quality score from 6/10 to 8/10.

Who is Y Combinator on YouTube?

Y Combinator is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

Ergonomic Mechanical Keyboard For Programmers Recommended

Provides tactile feedback and comfort for managing 10-15 parallel coding sessions and high-volume pull request workflows.

Amazon →

Ultrawide Monitor For Multitasking Productivity

Offers the screen real estate necessary to view multiple terminal windows and browser sessions simultaneously.

Amazon →

Vertical Laptop Stand For Desk Space

Helps organize a clean, minimalist workspace for managing multiple hardware or software development setups.

Amazon →

Book On Software Engineering Management Principles

Provides foundational knowledge on team structures and processes that the speaker adapts for AI agent management.

Amazon →

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Summarize another video

Full Transcript YouTube

Hi, I'm Gary, president and CEO of Y
Combinator. I'm also an engineer who
spent the first decade of my career
building software full-time. I studied
computer systems engineering at
Stanford, then was employee number 10 at
Palunteer, where I was an engineer,
designer, and product manager all at
once. I co-founded Posterous, a micro
blogging platform that sold to Twitter,
and I also built the first version of
Bookface, YC's internal social platform
and knowledge base. Basically, I've
written a lot of code in my career, and
I'm here to tell you we are in a
completely new era of building software,
the agent era. It turns out the way to
get agents to do real work is the same
way humans have always done it as a team
with roles with process with review. I
built GStack to encode this 3 weeks ago
and now it has more GitHub stars than
Ruby on Rails. In this video I want to
explain how it can help you build with
agents. I've coded more in the past two
months than I did in all of 2013, which
is the last time I worked really, really
hard as an engineer. I started playing
with Claude Code back in January after
hearing people like Andre Karpathy and
Boris Churnney say they weren't manually
writing any code anymore. And I got
completely hooked. Along the way, I've
essentially built all of Posterus, which
took 2 years to build with a co-founder
and a team of 10 engineers. I've
essentially built all of my startup
Posterus, which took 2 years, $10
million, and 10 engineers to build. Out
of the box, the model wanders. It
doesn't know your data well. So, it
guesses. And guessing at that scale is
how you get plausible looking code that
silently breaks. The bottleneck here is
not the model's intelligence. As long as
you set the models upright, they are
already smart enough to do extraordinary
work on your codebase. This is
backwards. The scaffolding should be
trivially thin. GStack is my
implementation of the thin harness fat
skills approach. It's an open-source
repo that I built that turns clawed code
into an AI engineering team for you.
skills that act like a team of
specialists. Office hours is one of
those skills. It's actually modeled
exactly after what we go through at YC
as a partner doing office hours with
startups. It starts by asking six
forcing questions for you to reframe
your product before you start building.
Let me show you how it works. The best
way to get started with GStack is uh
actually conductor. And so we're going
to go in quick start. And GStack is
actually built into Conductor right now.
You just click GStack. And today we're
going to make a tax app. Uh it's going
to go into your Gmail and fish out all
of your 1099s cuz it's tax day as of
today. Gstack is actually a set of
skills. And the first one that we're
actually going to use is called office
hours. This is actually the distilled
version of what is thousands and tens of
thousands of hours that the 16 YC
partners have spent many many years
honing and perfecting. And uh this is a
distilled down 10% strength version of
what we do at YC every day. So as you
can see uh conductor actually just drops
you right in there. We're in YC office
hours now and I'm trying to do a startup
to create
to help people get all their 1099 ins
out of
their Gmail and financial institutions.
Many banks will email you with uh new
tax documents, but some won't. So, we
need to both search the user's inbox and
accept URLs to go and search and
download the 1099 int PDFs.
Cool. That's our startup idea. It's just
uh something to help people with their
taxes and it's something that I had to
deal with just yesterday. So, the user
wants to do office hours about a startup
idea. um and it's starting on office
hours. So, one of the things you'll
notice is um we have Gary mode on and uh
what that does is it actually shows you
all of the reasoning traces. So, that's
one of the things I really like about um
using GStack. It actually with conductor
you actually get to see exactly what um
the model is thinking as it does it. Uh
right now it's just sort of getting
started but now it's starting to do some
context. Okay, this is a fresh project
with an initial commit. No prior design
docs. We're in startup mode. Oh, and
here this is uh this is sort of the
first thing that happens. You know, your
model with the skill uh of office hours
does a lot of thinking. It searches the
web sometimes and figures out well what
what are you trying to do? Here's the
question that determines everything
else. What's the strongest evidence that
you have that someone actually wants
this? This is actually one of the most
important questions to ask yourself when
you're trying to decide, should I work
on project X or work on that startup? I
actually just experienced this recently.
So, I'm going to click that. I have this
experience uh just the other day.
So, what happened exactly? How many bank
accounts do we have and which one sent
uh 1099? Well, I have a few and then the
process of hunting down is actually
quite painful. Okay, the pain is real,
but the consequence is friction and
annoyance, not actual penalties. Well, I
mean, frankly, my accountant uh sends me
annoyed emails, and sometimes they get a
phone call because they're just trying
to do my their job, and uh I'm supposed
to be there to help them.
And so, I do have more than five
accounts. This is actually a really
interesting thing. It points out,
Turboax and HR Block actually have 1099
import features, and Plaid connects to
banks. Why aren't those solving it for
you? So, I answered and it said, "Well,
the user's thinking bigger than a
document act creator. They're seeing
this as a funnel. The hook is we'll find
all your 1099 ins for you, solving an
immediate pain. But the expansion is now
that you have your docs, let's actually
get your taxes prepared, which is
matchmaking and lead genen for tax
preparers. And it's a classic wedge
strategy. It's actually a more
interesting business model than just the
document aggregation alone, which is
right. You could probably charge $2 or
$5 a month or $2 or $5 maybe an entire
year for $1099 in aggregation, but you
can actually charge um a percentage of
the transaction that ends up happening
with a tax preparer, which uh might be
10x more. This is actually a very cool
part of office hours. I might just build
this for myself just for fun because I
do have to do this every single year.
And this is actually one of the most fun
parts about using office hours and
GStack is it's not this sort of on rails
thing. It's more like a conversation
with your model. And so if you just came
in and gave instructions to the model as
it is, if I just typed the original
thing, which is I need to go and find my
1099s, it'll go do that. It will just
literally do it, but it won't think
about who's the user, what is this, what
is the business model, who wants this,
what's the pain point, how does it work,
right? Like this is the kind of stuff
that we get to do every day with lots
and lots of founders in office hours and
we're pretty good at it, but so is this
skill. Revised AI browser automation
GStack browser style lets the user log
in then AI navigates to tax docs and
downloads PDFs. The user watches it
happen. I think that's correct. Let's
see. I mean this is a very interesting
thing for me. I mean I feel like I learn
about the process of building a little
bit even though I do office hours with
people all the time. This is sharp push
back on premise 3. You're right. I do
have GStack browser and the approach
you're describing is genuinely clever.
The model would be user logs in AI takes
over, navigates to tax docs, finds the
1099 ins, downloads it. No plaid, no
stored credentials. The user watches the
whole thing happen in the visible
browser. I mean, what I like about it is
it can happen on their actual browser.
It's not happening in the cloud. The
cloud is just someone else's computer.
One of the cool things that GStack does
increasingly is it lets you use codecs
to actually sort out all of the crazy
bugs that might be in here. I'm not
going to do it right now, but it is
something that you can do when you're at
home working on it yourself. The way to
think about Claude code is that by
default it uses Claude. And I think Opus
4.6 is sort of ADHD CEO. He's the guy
you want to get a beer with and he's got
a billion ideas, but when the going gets
tough, you got to call in your autistic
CTO and that's Codex. All right, we're
going to skip for now because we're
actually pretty close. I feel like
basically we're in plan mode and office
hours helps us start off with a plan
that has a lot of the things thought
through. So, here's actually a really
cool uh example. It actually thinks
through and here's three different
approaches. The first approach is Gmail
off then search for tax doc not
notification then output a checklist of
banks which issue 1099s. There's no
browser automation initially. The effort
is small and the risk is small. You know
when I look at that it's like that
sounds interesting but it doesn't sound
big enough for me to actually even work
on this. Like I could do that myself.
Next is full stack Gmail and AI browser
automation using uh and a CPA
marketplace. This sounds like what I
want actually.
And then uh it sort of thinks out of the
box. It says, "Oh, okay. What about
approach C? CPA first flip the go to
market. You know, I would say B sounds
right." And then actually I sometimes I
like to add this extra thing which like
when I have an idea when I one of the
approaches speaks to me, but then I
think about something else. I'm like,
"Okay, well, I like B, but actually we
could use the browser
interaction
to skip Google OOTH entirely and just
have the user open Gmail and a version
of GStack browser could just use Gmail
to find the 1099s
to search for automatically."
Simultaneous
to that, it could also ask the user what
other banks they have. Also, and this is
what happens for me.
If they already have a CPA, you can find
out from the email. And if you're me,
you probably already have a bunch of
emails from your CPA bugging you for the
specific accounts.
We're sort of at the end of office
hours, but as you can see, we already
went from sort of a halfbaked rough idea
for something that we might want to do.
I'm not saying this is actually a good
startup idea, but you can see how this
got farther along. We started with
something that might start with OOTH and
then CPA's nagging emails, but in the
end we realized, well, we have a browser
and the browser could be used with
browser automation to search the inbox,
find all of the 1099s that you need to
download. It can also using LLM ask you
which bank portals you need to add to
and it can go log in with your account
and actually download the PDFs for you
and then send an email to the CPA. So I
really like this browser automation is a
very outofpocket sort of unusual way to
solve this problem. And the wild thing
about coding models is, you know, a year
ago, two years ago, even like 3 months
ago, it's not clear to me that anyone
would even try this. I think that's the
most interesting thing about uh our time
right now. You're able to have an idea
and then get farther along with it than
you ever would be. Frankly, sometimes I
use office hours and maybe one in three
times I get to the end of it and I say,
"You know what? This isn't something
that makes sense." You'll notice that
there's actually a feasibility aspect of
office hours and that's one thing I
really pride myself on in office hours
working with startups. I have a very
strong opinion about how the world works
and what might work and um it's just
very interesting to see Opus 4.6 mirror
that in trying to help you figure out
what your startup or product idea might
be. Now, what it's doing is a multi-step
adversarial review. It's trying to put
your idea through the paces. And as you
can see, it's already found a bunch of
things and it's going to try to autofix
it. There's no failure handling. There's
no privacy section. 2FA handoff has no
proposed solution. It actually tries to
auto fill out these things. And it if it
can, it does. And so our doc survived
two rounds of adversarial review. And it
automatically caught and fixed 16
issues. Um so we're going to approve
this design doc. So, as you can see, the
adversarial review improved the score
from 6 out of 10 to 8 out of 10 with
three remaining issues that we can worry
about later. Now that we've locked in
the adversarial review and addressed all
these issues, uh, normally what I would
do is run plan CEO review, but instead I
think what we're going to do is jump
directly to design shotgun, which is one
of my most fun uh, ways to use this. And
this is just one of a bunch of different
design tools that are in the bag. So, it
figured out here's a bunch of different
views. What do you want to actually
design? And let's just do the main
checklist dashboard. Design checks my
visual brainstorming tool. So, it'll
actually generate multiple AI versions
and then ask us questions about it.
These are three directions. It takes
about 60 seconds. it actually farms it
out to uh OpenAI codecs which um is able
to use image gen. So all right let's
there's three versions command center
friendly progress and split view. Let's
take a look. All right so let's let the
agents cook and we'll be back in about 5
minutes. Great. The agents are done
cooking. And this is what we we got
back. We got three different options for
the actual page that shows up in the
command center for tracking down our tax
documents. So let's look at them one by
one. There's option A, B, and C. All
right, here's one command center.
There's a dashboard. Here's all the
specific. I mean, this looks pretty
good. If you can extract here are all
the banks and here are all the 1099s and
where are they coming from? Um, and what
their status is, that's pretty good. I
like that. I'm I'm going to give that a
four out of five stars. Option B is like
much more friendly. Um, so option A is
sort of like if you're a Linux hacker, I
bet you would really like this. But
option B, I think it's more friendly for
just normal people. So I kind of like I
I I might put that as a five. That might
be a pick. And then let's see. Option C.
This makes it way more complicated than
it needs to. So I really wouldn't do
that. Let's go with option B.
Uh, and then the cool thing is if you
don't like it, you can enter uh, you
know, any of your feedback. You can
click regenerate. But in this case,
we're just going to run with option B
and continue. So that comes back in. And
you know, as you can see, we're going to
go ahead and select option B. And there
it is. So a friendly card-based approach
with progress and the progress ring.
Good instinct. Variant B is locked in.
So while I have you, I mean, that is
just two of 28 different commands. We've
got more than 70,000 uh stars now. And
some of the people who use it like they
they actually talk about how when
they're using cloud code they spend 80
to 90% of their time in office hours
plan CEO review and auto plan. Um this
is sort of a rough view of how that
sprint process actually works. We
already talked about office space, but
if you don't want to do a lot of back
and forth, if you don't want to be in
the weeds, I did create auto plan, which
gets you through CEO, engineering,
design, and developer experience review
using basically my default
recommendations. Like these are sort of
programmed to be what I would do if I
were you. There are a bunch of design
skills that you can use after the code
is actually done. Cloud code will
actually build when you click approve on
the plan and then after it's done
writing the code you can run review
which does a staff level uh bug catching
service that goes through puts the work
through the paces full code review uh
finding bugs that might not have been in
the plan mode and then the coolest part
I think that um is actually an
incredible amount of code is I wrote a
CLI around playrite and chromium
So there's actually an entire headed and
headless browser in there. And that was
a real magic moment for me as I was
using cloud code as I sped up. Um
there's this idea of trying to get a to
a level 8 software factory and GSAC does
not get you to level 8, but I do think
it gets you to level seven. And that's
where I can run multiple conductor
windows on different projects and
sometimes three or four all on the same
project all at the same time. These are
parallel PRs with parallel branches and
parallel different features that all can
land more or less simultaneously. And
one of the bottlenecks I ran into was
that, you know, once the agent was doing
all the work of planning and design and
coding it, I found myself sitting there
doing QA, probably the least fun part of
software development. So that made it
very, very important for me to try to
automate that. And when I did, Claude in
Chrome MCP is one of the worst pieces of
software I've ever used. You know, every
time it would try to do an action, it
would think and think and think. There
was crazy context bloat. Often it
wouldn't even do anything, but it would
take two to three seconds even when it
was working to be able to take an
action. And I was amazed that I could
use all of my other skills in GStack to
create the SLQA and SL browse tool. I
basically wrapped Playright at the CLI
level. And now your cloud code and any
agent now can actually just use the
browser. And so you know, not only could
it use the browser, it can take
screenshots. It can do complex
interactions. It can click on things. It
can fill things out. Now it can even
download media, run eventually full
regression tests and update CSS and
assess real browser bug issues, whether
it's JavaScript or CSS. And finally,
there's a ship tool. So, it's sort of
the last step before to make sure that
your PR is ready to land on main. And
this is actually how I work. I run 10 to
15 parallel cla code sessions all at the
same time. I might in one session be
running office hours on a brand new
idea. And I actually now have multiple
open- source projects with tens of
thousands of stars. And I'm probably
sitting on about 400 uh PRs to review
right now. And so I almost always have
one or two sessions active for each
project just evaluating and bringing in
all the open- source fixes that I'm
getting from the community. Uh and I
evaluate it in waves. Um one of the
things that's been really scary in AI
coding right now is supply chain
attacks. So I'm really really paranoid
about it. But the great thing is I have
GStack that has my back. So I don't have
a to-do list anymore. One of the things
that has emerged is I actually click on
whenever I have an idea or I get a bug
report from a user or I see something on
X where someone's frustrated with what
GStack or GBrain does, I just click the
plus icon in Conductor. It creates a new
work tree and each one of these things
is a new work item. And all I have to do
is run office hours, CEO review, end
review, uh adversarial review, and then
I just run my normal process. when it's
ready to land, it lands and I can do 10,
15, 20, sometimes 50 PRs in any given
day, depending on the number of meetings
I have in that day. So that's it. Uh,
GStack is available right now. Just go
to github.com/gritan/GStack.
When you run/off hours, you're getting a
version of the real product thinking we
do at YC with founders. similar push
back and similar reframing before you
ever meet us. Give it a try and let me
know what you think. This is the most
incredible time in history to build
software. The barrier to building just
collapsed. The only question left is
what are you going to build? It's time
to let it rip. Go make something people
want.

Help & FAQ

Beyond Bigger Models: Recursion As The Next Scaling Law In AI

Y Combinator

May 01, 2026

The GStack Framework

Workflow Demonstration: Building a Tax App

Scaling and Automation

Managing Parallel Development

Takeaways

Frequently Asked Questions

What is the 'Office Hours' skill in GStack and how does it shape product ideas?

How does GStack’s adversarial review improve a design document?

Who is Y Combinator on YouTube?

Does this page include the full transcript of the video?

Helpful resources related to this video

Share This Summary

Embed This Summary