00:00:00The CEO of YCombinator has built his own toolkit for Claude's code called GStack, his secret
00:00:06to crushing almost a hundred PRs in seven days, which includes nine specialised workflows,
00:00:13a headless browsing mode using Playwright, Greptile integration, a diff-aware QA and much
00:00:18much more.
00:00:19But Gary's recent tweet about the future of code has got a lot of developers really
00:00:24annoyed.
00:00:25So what does that mean for the future of GStack?
00:00:28Subscribe and let's find out.
00:00:33Gary Tan has been the CEO of YCombinator since 2023 and before that he co-founded a venture
00:00:39capitalist firm in 2011.
00:00:42So he has loads of experience when it comes to going through pitches and finding out what
00:00:46makes a new piece of tech unique.
00:00:49And he's put all of that knowledge into his own Claude's code toolkit, which you can
00:00:53tell by looking at the names he's given to a lot of his workflows.
00:00:57In fact, let's give GStack a spin.
00:00:59So for GStack to work, you'll need to have Claude code installed as well as BAN, but once
00:01:03you've properly installed it on Claude's code by prompting with this exact text or from
00:01:08just downloading the skills, you should have this information added to your Claude MD file.
00:01:12Mine was empty.
00:01:13That's why this is the only thing here, but if you have some text, then this will be added
00:01:17to it.
00:01:18It also puts all the relevant skills into the skills directory if you want to share it with
00:01:21your teammates and then installs Playwright with the appropriate browser.
00:01:25Now I'm going to use GStack to add a feature to this React Vite application to give the
00:01:30user the ability to download an image of a tweet with a specific URL.
00:01:34Now you may have seen me add this feature in a previous video.
00:01:37I'll have a link to it in the description if you want to see what the results were, but
00:01:41we'll see if GStack can do better than that.
00:01:44So first I'll need to start in plan mode, then use the plan CEO review skill and give GStack
00:01:49some information about the feature.
00:01:51Now I'm going to say add a feature that takes a screenshot of a tweet from the URL provided
00:01:56by the user.
00:01:57I also want the user to customise and download the image and I want Claude to honour the existing
00:02:02layout and styles.
00:02:03So after I hit enter, GStack first checks if there are any updates to that skill and then
00:02:08checks the git log before proceeding.
00:02:10Now this mode rethinks the problem from the perspective of a founder/CEO and tries to think
00:02:16of the best possible version of what we're trying to build and challenges assumptions
00:02:20about the scope and value.
00:02:21So once it's done that, it allows us to choose how much we want to challenge the original
00:02:26scope.
00:02:27And here I'm going to go with the scope expansion because it has the most amount of features.
00:02:30Then it lets us choose a critical architectural decision.
00:02:33I'm going to go with the recommended since it's the easiest.
00:02:36And then it asks a few more questions, which again, I'm going to go with the recommended
00:02:39approach.
00:02:40And now that it's finished, it's come up with a mega plan showing the scope mode selected
00:02:44and everything it's going to do that's in that scope.
00:02:47And it's also written some things that are not in scope for this feature.
00:02:50And then down here we have the implementation plan, which has an architecture diagram, key
00:02:55decisions and different steps.
00:02:57This is an insanely detailed plan similar to something I'd get from superpowers if I went
00:03:01through the same route.
00:03:02Note, there's also a plan engineering review skill in GStack, which turns Claude into an
00:03:07engineering manager or tech lead to come up with architectural diagrams, lock in the tech
00:03:12stack, define edge cases and so on.
00:03:15But it looks like the plan CEO review skill has got ahead and done some of that already.
00:03:20So we're jumping to the implementation.
00:03:22And now that it's done, we can run the review slash command to review missing edge cases,
00:03:27find bugs that would have passed CI and basically catch any issues before they hit production.
00:03:32So again, that checks for new updates inside the script, checks the diff.
00:03:36And now it's checking the completeness of the task before giving us a summary saying that
00:03:40no issues have been found.
00:03:41And now we can run the ship slash command, which syncs with the main branch, runs tests
00:03:46and resolves any grep tile reviews if they exist.
00:03:49And here we can see it's gone ahead and created a pull request without me even telling it to.
00:03:54And then at this stage we can run the QA slash command, which will test only the changes we've
00:03:58made based on the diff.
00:03:59And here we can see it started the server locally, and it's going through the website to test
00:04:05the features that have just been implemented using screenshots and much more.
00:04:09It's found some 500 errors from screenshots and has found a bug with JSON pass, which it
00:04:15looks like it's fixed.
00:04:16Here we go.
00:04:17It's verified and pushed the fix.
00:04:20And now it's written a final report with the issues that it solved.
00:04:24This is very cool.
00:04:25Okay.
00:04:26So now it's done.
00:04:27Let's go ahead and try the feature.
00:04:28And now we have a screenshot page.
00:04:30Let's grab a tweet from Tana.
00:04:32So this one, and I'll paste that in here.
00:04:34It's not the most exciting tweet, but it's just a test if this works.
00:04:37And wow, okay, this is super impressive.
00:04:40We have the tweet here.
00:04:42We can pick between lights and it's capturing again.
00:04:44Oh, wow.
00:04:45Okay.
00:04:46So we've got light and dark mode.
00:04:47We'll see if it's cache that.
00:04:49And it has very cool.
00:04:51I can hide the actions and here we go.
00:04:53So I can show and hide the images and I can change the background.
00:04:58This is very cool.
00:04:59So you've got LinkedIn, we've got Twitter, blog, gradient purple, and we can even customize
00:05:03it or change the angle of the gradient.
00:05:07Wow.
00:05:08This is super fully fledged and we can change the aspect ratio.
00:05:11So we've got nine by 16, 16 by nine, one by one and so on.
00:05:16Let's now actually downloads the image.
00:05:18And here we go.
00:05:19If I now click on this, you've seen all my tabs.
00:05:22We have the image here.
00:05:23I'm going to open it and preview.
00:05:24And this is it.
00:05:25This is the image I just took with the feature I just built with G stack, which is insanely
00:05:29impressive, but there's more that we can do because if we go back to the PR, we can see
00:05:34Greptile has a summary, so it's found some resource exhaustion from the server, race condition,
00:05:40no cash expiry, and so on.
00:05:42And instead of me asking Claude to look at the issues and solve them, we're just going
00:05:47to run the review slash command.
00:05:49It's found all the comments.
00:05:50It's given me some options down here on how to fix them, which I'll go through.
00:05:53And now it's fixed all the issues.
00:05:55Well, apart from one false positive and has pushed the code, Greptile seems to be happy.
00:06:00As someone who regularly uses superpowers, I can already see the benefit of G stack, even
00:06:05though some aspects of it are quite complex.
00:06:08But what about Gary's comment on Twitter saying that Markdown is the new code?
00:06:13Well, I can kind of see where he's coming from.
00:06:15I don't think he's saying someone with a computer science degree has wasted their time purely
00:06:20because you can write Markdown and it will write the code.
00:06:22I think it's more to do with the instructions because newer models are getting better at
00:06:27obeying Markdown instructions before there was a time when I would need to have a Claude
00:06:32code hook just to make sure he uses Bun to install instead of using NPM.
00:06:36But now I can put it in the Claude MD and with a good model like Opus, it tends to obey 90
00:06:42to 95% of the time.
00:06:44So I think what he's trying to say is that if you have a detailed enough and well-structured
00:06:49Markdown file, the model can create a good piece of software based on those instructions.
00:06:55But this isn't to say that GStack is just a bunch of Markdown instructions.
00:06:59Each skill has its own directory, even the ability to upgrade GStack.
00:07:03And if we focus on the browse skill, we can see there's a template file and the actual
00:07:08skill file.
00:07:09And this isn't anything to do with Go templates, regardless of what the GitHub page says.
00:07:14The way this works is if we go to scripts and then we go to genskill, the TypeScript file
00:07:20will read the template files and replace any placeholders inside them with actual Markdown.
00:07:26But I'm not going to focus on each skill individually because they're quite detailed.
00:07:30But what I will focus on is that the browse skill has more than just a skill MD file because
00:07:35we have a test directory here and we also have the source directory which contains the actual
00:07:40implementation for browser management and so on.
00:07:42So we can already see that the commands here are fairly involved.
00:07:46But if we take a look at the changelog, this shows some really interesting features like
00:07:49end-to-end observability, incremental eval saves and so on, which is used for developing
00:07:55the app.
00:07:56It shares reviews in a to-do format.
00:07:58It supports screenshot element and region clipping, not to mention all of the integrations it has
00:08:03with Greptile and the fact that it was built with Conductor in mind.
00:08:07So the million dollar question is, will I personally use GStack?
00:08:11And I would say actually I'm going to try it out for 30 days.
00:08:15So I'm going to delete the superpowers plugin and make GStack my main code tool for preparing
00:08:21features and fixing bugs and see how it goes.
00:08:23Who knows?
00:08:24I might just clone the next Vercel open source tool and start some more beef on Twitter.