First, I appreciate how he implemented auto-update. Not sure if that pattern is original, but I've been solving it in a different-but-worse way for a similar project. NOT a fan of how it's being used to present articles on Garry's List. I like the site, but that's a totally different lane.

The skills are great for upleveling plans. Claude in particular has a way of generating plans with huge blind spots. I've learned to pay close attention to plans to avoid getting burned, and the plan skills do a fair job at helping catch gaps so I don't have to ralph-wiggum later. I don't find the CEO skill terribly effective, but I do like the role it plays at finding delighters for features. This is also where I think my original prompting tends to be strong, which could be why it doesn't appear to have a huge impact like the other skills.

I think the design skills are great and I like the direction they're going. DESIGN.md needs to become a standard practice. I think it's done a great job at helping with design consistency and building UIs that don't feel like slop. This general approach will probably challenge lots of design-focused coding tools.

The approach to using the browser is superior to Claude's built-in extension in pretty much every way (except cookie management). It's worth it for that alone.

For people who don't understand this...think of each skill like a phase of the SDLC. The actual content, over time, will probably become bespoke to how your team builds software, but the steps themselves are all pretty much the same. All of this is still early days, so YMMV using these specific skills, but I like the philosophy.

If he weren't the CEO of YC, this wouldn't be on PH, and it wouldn't be on HN.

This is not an impressive setup, folks. It's overengineered and deeply into its own form -- it will not make your agents better, and is likely to make it worse. There are lots of other people to follow/learn from/mimic for skills/context engineering.

My guesses would be five digits and 90%.

But maybe there is some cool stuff here. A lot of prolific AI-assisted engineers I know have their own advanced plan modes, and the CEO plan mode in the repo is interesting (although very token heavy)

https://github.com/garrytan/gstack/blob/main/plan-ceo-review...

It is always a learning exercise to see how other people are using CC and I'm sure I'll learn a lot from this, so thanks for sharing it.

But, I don't understand what 600,000 lines in 60 days mean. Lines of code is one thing, but to do what? There still needs to be a loop where CC generates code, there is test automation, maybe do some code review, and then test/run to see what it's built and if it matches the spec, refine the spec, provide new guidance and so on. Products are not built in isolation and are not just KLoC.

Now, if I were asking CC to, take the Algorithms text book and write all the code in all the language etc. (as an example) 600 KLoC over 60 days would make sense. If it were porting an existing product from one stack to another, maybe. But for new products, at least to me that part doesn't make sense.

Now I know these are symptoms of bipolar disorder/psychosis (they both eventually got professional treatment and told me much later), and I wish I’d known at the time so I could’ve helped. He’s bragging about sleeping 4 hours and joking about having cyber psychosis. [0]

Sleeping only 4 hours is a classic mania symptom.

I’m not as close to Garry, so I don’t know for sure, but some of the behavior feels very similar to what I’ve seen in my friends.

I hope Garry has people in his life who can help. At the very least, you have to sleep — poor sleep is strongly correlated with psychiatric conditions.

[0] https://youtu.be/W3YpC4Dvzso?t=929

edit: There's a few funny threads on other social media. Honestly, though, let a guy get excited, when you find new ways of using new tech; he's one of the lucky 10,000 who has discovered prompt scaffolds. There are better, bespoke tools for more targeted tasks.

LOC will never be a good metric of software engineering. Why do we keep accepting this?

I can generate 1 million LOC if I really wanted to.

As long as LOC is the main metric for these setups, they will never be successful.

That would be considered a huge liability and shameful historically.

They’ve recently started using their AI pipeline to put out rust-based conversions of tools and it seems to be going pretty well.

No you haven't.

and what is there to show for it? absolutely terrible metric

https://www.youtube.com/watch?v=Q6nem-F8AG8

The way the whole repo is written it’s like he thinks he is the messAIah. We are all getting sold glass marbles.

Get some damn sleep Gary.

What a disgrace, hacker culture died to this

Is any of it trustworthy?

Not to mention using lines of code as a metric of usability is just _whatever_.