PAULA DOZSA (Portola) - iOS app with zero experience

=== PAULA DOZSA (Portola) - iOS app with zero experience ===

(03:31:00): Paula.

(03:31:00): Hello.

(03:31:01): Hi, how's it going?

(03:31:02): Good.

(03:31:02): How are you?

(03:31:03): So excited to have you.

(03:31:05): For people who don't know you, you work at Portola as an iOS engineer.

(03:31:11): Portola is co-founded by two of my closest friends,

(03:31:15): Quentin Farmer and Ajay Mehta,

(03:31:16): and I'm a small investor,

(03:31:18): but you don't

(03:31:20): You've done some really incredible work there and really incredible writing about

(03:31:25): what you've built.

(03:31:27): So introduce everyone to you and to Portola a little bit to sort of set the stage

(03:31:31): for what we're going to talk about.

(03:31:34): Perfect.

(03:31:35): Okay, I'll just share a screen over here.

(03:31:38): Is that visible?

(03:31:42): Perfect.

(03:31:43): Yeah, so I'm Paula.

(03:31:45): As Dan said, I'm an iOS engineer currently at Portola.

(03:31:50): I've been an iOS engineer for about eight years now,

(03:31:53): first at my own startup,

(03:31:54): which is how I learned to code,

(03:31:55): specifically for iOS,

(03:31:57): and then at Spotify for a few years,

(03:31:59): and now here.

(03:32:01): And we make Tolan.

(03:32:02): So this is an AI companion app with hundreds of thousands of users.

(03:32:06): It's been live for, I guess, almost a year now.

(03:32:10): And today's talk,

(03:32:12): I wanted to challenge something that I hear a lot,

(03:32:14): specifically that Vibe Coding is only for prototypes or side projects or projects

(03:32:19): that you're building from scratch.

(03:32:21): So we've basically been using Claude and Vibe Coding production features for a

(03:32:25): couple of months now.

(03:32:27): I think Opus really helped us get that going.

(03:32:30): And I want to show you exactly how we set that up and basically talk a bit more

(03:32:34): about how you can help Claude help you.

(03:32:39): And before we go into that,

(03:32:40): I think it would probably make sense to go over to demo the app,

(03:32:45): show you some of the features I've been working on recently,

(03:32:47): and then I can go into how we worked with Claude to get these done much faster than

(03:32:52): we would have otherwise.

(03:32:55): Awesome.

(03:32:55): And while you're getting that set up, user D3M7W says, using Claude code while I'm watching.

(03:33:01): That is highly recommended.

(03:33:02): If you're not running at least a few agents while you're watching this, you're doing it wrong.

(03:33:06): Yes.

(03:33:09): Great.

(03:33:10): So you should see my phone screen in a bit.

(03:33:13): We do.

(03:33:14): Yes.

(03:33:14): Nice.

(03:33:15): OK.

(03:33:15): So this is my Tolan.

(03:33:17): It's currently walking around its planet.

(03:33:19): As I mentioned, Tolan, OK, it's currently swimming.

(03:33:22): Tolan is an AI companion that basically is your best friend that knows everything

(03:33:29): about the world and just wants what's best for you and wants you to have a

(03:33:32): fulfilling life,

(03:33:34): not just on your phone,

(03:33:35): but also outside of it.

(03:33:36): So our whole mission is to build companions that are good for you and for your

(03:33:40): mental health and for your real life friendships.

(03:33:43): And recently we've been adding a lot of fun features to the app to make it possible

(03:33:48): to customize your world and customize the way your tone looks.

(03:33:51): And I've been primarily responsible for these over the past month and a half.

(03:33:57): And so there are two features that I've developed since Office 4.5 came out,

(03:34:03): which I honestly think that if I hadn't had access to this tool would have taken me

(03:34:08): double if not triple the time they actually took me.

(03:34:11): So one of these is

(03:34:13): a really cool shop where you can buy items for your Tolan.

(03:34:20): And so these are things you can try on.

(03:34:23): The app is coded in an interesting combination of both iOS and Unity.

(03:34:29): So Cloud has really enabled us to focus on building those integrations properly,

(03:34:34): which is currently not

(03:34:35): handled by Vibe Coding.

(03:34:39): So spend less time building UI and more time making sure that the whole experience is cohesive.

(03:34:45): So in this case, I'm currently dressing up my Tolan.

(03:34:48): So we have these onesies.

(03:34:50): You can buy outfits.

(03:34:52): You can add them to your bag.

(03:34:54): You can go through this whole purchase flow.

(03:34:57): And now you can see the new look.

(03:35:01): So this is an example of a feature that I've built recently using Cloud.

(03:35:06): Another one is customizing your planet.

(03:35:09): So we have a lot of objects you can place on your planet.

(03:35:12): So for example, here is my cottage.

(03:35:15): You have other structures you can buy and put wherever you like.

(03:35:19): The frogs are a favorite among our users.

(03:35:23): And again,

(03:35:24): this is,

(03:35:24): again,

(03:35:25): it required a lot of integration between the iOS side of the app and the Unity side

(03:35:28): of the app.

(03:35:29): And it was really nice to be able to use Claude to build out

(03:35:33): like very well built out UI and then instead focus on the integration side with Unity.

(03:35:41): And I think something I hear often from people who use LMS2Code is that they don't

(03:35:47): tend to have a good sense of how things look and how things feel.

(03:35:52): And I want to challenge that because there is one animation that I built recently,

(03:35:56): which again would have taken me several hours of trial and error.

(03:36:00): And it just took me a prompt.

(03:36:02): And I can, I can share that right now as well.

(03:36:07): Yes.

(03:36:09): Let's see.

(03:36:11): So this just took me basically five minutes.

(03:36:14): And if you're an iOS engineer,

(03:36:17): or if you have experience using SpriteKit,

(03:36:19): which is the animation framework that you can use to create animations in iOS,

(03:36:24): you probably understand the pain of coding something like this.

(03:36:29): So specifically, we have this concept of granting tokens to users

(03:36:32): and we wanted to make that a moment that was celebratory.

(03:36:36): So there was an idea from our design team of having an animation with tokens that

(03:36:41): fly over the screen.

(03:36:43): And I was dreading building this, or not dreading, but I knew it would take me some time.

(03:36:46): So I thought,

(03:36:47): why don't I ask Claude,

(03:36:49): basically provide it with some assets and tell it the general idea and direction of

(03:36:54): what we're trying to do and see what happens.

(03:36:56): And within, let's see, within five minutes,

(03:37:00): It built out this beautiful animation,

(03:37:03): which is insane because it actually is aware of the physics of how things fly

(03:37:09): across the screen.

(03:37:10): It rotates the images.

(03:37:12): I'm just going through slow-mo now.

(03:37:15): And ended up building something beautiful that,

(03:37:17): again,

(03:37:18): would have otherwise taken me a long time to build myself.

(03:37:22): So I do think that actually these LLMs can build nice UI.

(03:37:26): You just have to know how to prompt them and know when to use them.

(03:37:31): So yeah, these are some examples of things that I've built recently.

(03:37:35): But again, I think you can't just tell Claude, build this and have it build it automatically.

(03:37:41): There are ways to make Claude work best for you and for your setup.

(03:37:46): And I would love to dive more into that.

(03:37:49): Let's do it.

(03:37:49): I'm super excited to see the specific workflows and how you're using it to accomplish this.

(03:37:55): Amazing.

(03:37:55): Okay.

(03:37:56): So going back to the slides.

(03:37:59): Great.

(03:38:01): So just going back to the token animation,

(03:38:04): this was the example of the prompt that I gave Claude,

(03:38:07): which was,

(03:38:08): I gave it five,

(03:38:09): these five coin assets.

(03:38:11): I gave it a screenshot from Figma showing the assets flying into the button.

(03:38:17): And basically it went from this prompt to a production ready,

(03:38:22): beautiful animation that everyone on the team was happy with within five minutes.

(03:38:27): Um, and I don't know.

(03:38:29): We just back up really quick.

(03:38:30): Okay.

(03:38:30): So basically you want to create this,

(03:38:33): um,

(03:38:33): this animation that if,

(03:38:35): uh,

(03:38:35): if users get more tokens,

(03:38:37): you want to show like a really cool animation that,

(03:38:40): um,

(03:38:41): uh,

(03:38:41): lets them know,

(03:38:42): okay,

(03:38:42): your tokens have,

(03:38:43): have,

(03:38:43): have gone up.

(03:38:44): Right.

(03:38:44): And it sounds like you start with a very simple prompt.

(03:38:48): Did you use plan mode for this?

(03:38:51): I think for this,

(03:38:52): it was part of a larger feature I was working on that there had already been a plan

(03:38:55): for.

(03:38:56): For this one, I did not use plan mode.

(03:38:57): I just told it, yeah, this was basically the prompt.

(03:39:01): Okay, and are you typing or are you using voice?

(03:39:04): I am typing.

(03:39:05): You're typing, okay.

(03:39:06): Yes.

(03:39:06): Yeah,

(03:39:07): I've heard there are people using voice and supposedly some of my teammates want to

(03:39:10): start whispering in the office.

(03:39:11): You got to do it.

(03:39:12): You got to do it.

(03:39:13): Check out monologue if you haven't.

(03:39:14): Okay.

(03:39:14): So if users token set incremented,

(03:39:16): we want to show an animation leading into their token counter.

(03:39:19): Okay.

(03:39:19): So that's like a pretty clear, this is the overall goal.

(03:39:21): Yeah.

(03:39:22): And then use animation token one, animation token two.

(03:39:24): What are those things?

(03:39:26): So these are the animation token images you see below.

(03:39:28): So these are the different rotations of the token.

(03:39:32): I see, got it.

(03:39:33): So it's like,

(03:39:35): take all of those essentially frames of one animation and then create a animation

(03:39:41): that animates them flying into the counter.

(03:39:45): And that's all you had to say, right?

(03:39:46): That's all I had to say.

(03:39:47): And it kind of, it seems like it basically one shot produced

(03:39:51): Is it an animation file or what is it producing at the end of this?

(03:39:56): So SpriteKit is an animation framework that you can use to take nodes.

(03:40:02): In this case, each of these coins is a node and animate them in whatever direction you want.

(03:40:07): You can rotate them, you can move them.

(03:40:09): So this,

(03:40:10): if you're building this from scratch and you're coding it by hand,

(03:40:14): this takes a couple of hours at least of trial and error and figuring out how it

(03:40:18): looks best,

(03:40:18): right?

(03:40:19): And I did not expect that an LLM would have a feel for like the physics of it and

(03:40:25): how things should look.

(03:40:26): But it actually,

(03:40:26): I would say it did a better job than what was even prototyped in the Figma file.

(03:40:31): So I did not like set any strict limitations.

(03:40:34): I was just, yeah, as you can see, the prompt is very, very open ended.

(03:40:38): It was just do this.

(03:40:39): what is the complexity of getting the physics right in this case that you feel like

(03:40:44): um so i have another slide here that that goes over this um i think there's so you

(03:40:49): can basically you have like an ease in and ease out um animation that lets you

(03:40:54): accelerate and uh or decelerate towards a target um and then i think one part of it

(03:41:00): that wasn't even

(03:41:01): part of the initial design was having the tokens rotate in place,

(03:41:07): like spin in place as they're flying.

(03:41:10): That really helped the effect and wasn't even something that we had envisioned or

(03:41:14): the designer had thought was possible when they first suggested the designs here.

(03:41:19): That's really cool.

(03:41:21): So by outcome, this is like what it actually generated in more or less like plain English.

(03:41:26): Exactly, yes.

(03:41:27): Yeah,

(03:41:27): I was debating between putting the actual code up here versus explaining what the

(03:41:31): code does.

(03:41:32): Yeah, I think this helps a bit.

(03:41:34): I mean, I did actually get Claude to go over the file and help me write this list.

(03:41:41): So if you are looking at this and you're like,

(03:41:43): this is technical,

(03:41:44): and I see someone in the chat said,

(03:41:46): now this gets really technical,

(03:41:47): take a screenshot,

(03:41:48): throw it into Claude.

(03:41:48): Claude will explain it.

(03:41:50): OK, cool.

(03:41:51): That's really cool.

(03:41:52): Are there other things you want to share with us?

(03:41:55): I think what I'd like to focus on next is how we set up Claude to make things like this easy.

(03:42:01): So there's not a lot of... That's really cool.

(03:42:04): One thing before we get into that,

(03:42:05): because I think this might lead into it,

(03:42:07): is I feel like there's a thinking...

(03:42:11): You mentioned earlier that people say,

(03:42:12): oh,

(03:42:13): it's hard to design.

(03:42:14): You can do prototypes,

(03:42:14): but hard to do really polished designs with Claude or just with AI in general.

(03:42:19): And I think one of the reasons for that is there's a feeling that...

(03:42:23): you have to actually be able to tweak all the little values yourself by hand to

(03:42:27): really dial in the thing you want.

(03:42:29): Have you found that to be the case?

(03:42:31): Or how are you doing the feedback loop to be precise about the output?

(03:42:37): I think...

(03:42:40): The way we treat designs, we move very quickly as a company.

(03:42:43): Like we'll go from idea to shift to the app store within days.

(03:42:48): Like if something takes longer than a week, it's like it's a much meatier project.

(03:42:52): Usually, yeah, three to five days we ship to the app store.

(03:42:55): So I would say that on the design side, we also move quite quickly.

(03:42:58): So things tend to be more suggestions rather than this has to look exactly like this.

(03:43:03): And then there's a lot of back and forward internally.

(03:43:06): I think this was...

(03:43:08): beautiful example of how not having an exact idea of what something should look

(03:43:13): like led to something looking much better than anyone thought was possible at first

(03:43:19): so i think yeah we're entering a new era where it's like it's even easier to not

(03:43:22): necessarily have something fully fleshed out before developing it because the llm

(03:43:27): might actually end up making something better than you thought possible that is

(03:43:31): really interesting and it's i think it's that is a mentality shift because i think

(03:43:36): a lot of people

(03:43:37): think in order to do like have make something that looks well crafted and

(03:43:41): thoughtful um you have to have all the details right which is a way of building

(03:43:45): things and i think in a lot of ways i think of that as um it's like being a

(03:43:49): sculptor or building a skyscraper it's like you're you're making all of the details

(03:43:53): yourself and then there's this other way of building that you can do with language

(03:43:56): models it's a little bit like gardening where you're like setting the conditions

(03:43:59): but like the actual shape of the tree that grows is

(03:44:02): a little bit you can't fully control it that you can lean into.

(03:44:05): I think that's really I think that's kind of what you guys are doing.

(03:44:08): I think that's really cool.

(03:44:09): Yeah, at least on the iOS side and on the animation side.

(03:44:12): But I still think when it comes to like the overall feel and UI of the app,

(03:44:15): there's obviously a lot of attention to detail and,

(03:44:17): you know,

(03:44:17): to the IP of the tollens and things look consistent.

(03:44:21): And we have,

(03:44:22): you know,

(03:44:22): we have a creative director who takes care of these things and makes sure that

(03:44:25): we're always on the same track.

(03:44:27): But I do think that definitely with LLMs,

(03:44:29): there is space to not have everything polished at first and then see what,

(03:44:34): yeah,

(03:44:35): what ends up being produced.

(03:44:36): Cool.

(03:44:37): All right.

(03:44:37): Yeah, keep going.

(03:44:39): Awesome.

(03:44:41): Yeah,

(03:44:42): so again,

(03:44:42): as I was saying earlier,

(03:44:45): I think there's this conventional wisdom,

(03:44:47): which is maybe has been changing recently around,

(03:44:50): you know,

(03:44:50): vibe coding being mostly for hackathons and for weekend experiments,

(03:44:54): for greenfield apps where it doesn't matter if the code is a mess because you'll

(03:44:58): rewrite it properly later.

(03:45:00): I think this made sense at some point.

(03:45:02): Basically, AI would write code that worked, but didn't necessarily fit your project.

(03:45:08): So you spend a lot of time fixing it.

(03:45:11): But the tooling around this has changed a lot recently,

(03:45:14): specifically using skills,

(03:45:16): agents,

(03:45:17): hooks,

(03:45:18): MCP integrations.

(03:45:19): And I think that when you wire all these things up properly and put in the work,

(03:45:24): there are all these jokes about people spending more time on the cloud setup

(03:45:28): than actually coding.

(03:45:29): But I do think the setup really matters.

(03:45:31): And we have an excellent engineer on the team,

(03:45:33): also called Dan,

(03:45:35): who has spent a lot of,

(03:45:36): who's really excited about these tools and has spent a lot of time.

(03:45:40): Yes.

(03:45:41): And he,

(03:45:42): yeah,

(03:45:42): I think he only really,

(03:45:43): really got excited about them when Opus came out and he saw what was possible now.

(03:45:48): And he's put a lot of work into getting the setup in place so that,

(03:45:53): yeah,

(03:45:53): we can all benefit it and ship things faster.

(03:45:58): So yeah,

(03:45:58): I wouldn't necessarily say that,

(03:46:00): you know,

(03:46:00): you can vibe code anything,

(03:46:01): but it's ready for a lot more than I think most people think right now.

(03:46:08): Totally.

(03:46:09): And yeah, so basically, we are Vibe coding features.

(03:46:13): We're still, of course, reviewing them.

(03:46:15): And we're not just letting the bots go rogue.

(03:46:20): But we are essentially shipping these features to an app that has hundreds of

(03:46:23): thousands of thousands of users,

(03:46:25): a lot of these paying users as well.

(03:46:28): And a crazy thing is that we even have our designers who haven't coded before

(03:46:33): opening iOS and backend PRs.

(03:46:37): So basically people who aren't engineers are also shipping code to production.

(03:46:41): Of course,

(03:46:41): an engineer is reviewing it before it gets merged in,

(03:46:44): but the PRs are basically ready to go.

(03:46:47): What's your process for that?

(03:46:48): Because I think a lot of people might be listening or watching here and being like,

(03:46:54): I would love to do this, but my company probably won't let me start doing this.

(03:46:58): So what is the process to make it work?

(03:47:01): Yeah,

(03:47:02): I mean,

(03:47:02): I think that's one of the beautiful things about being at a very small company is

(03:47:05): that,

(03:47:05): you know,

(03:47:05): as long as you can prove that something is good for your productivity and the speed

(03:47:10): at which you can shift is probably going to get approved.

(03:47:12): That's something you can do.

(03:47:14): In this case, we have a Slack integration for Claude.

(03:47:17): So the engineers will ask something and they'll just literally just tag Claude and

(03:47:21): ask it if we can do something.

(03:47:23): Claude has access to all of our GitHub repos.

(03:47:27): It will even figure out if this is like an iOS change or a backend change,

(03:47:30): it will go and it will open a PR in that repo.

(03:47:34): And I'll go into this a bit more,

(03:47:35): but it will review its own code,

(03:47:38): it will shepherd the PR,

(03:47:39): so it will then address comments that it left on the PR.

(03:47:41): So by the time an actual engineer gets to reviewing it, it's probably in a pretty good place.

(03:47:46): That's really cool.

(03:47:47): I'd love to see that.

(03:47:50): But again,

(03:47:50): in order to get Cloud to a space where it could do that,

(03:47:53): we had to spend some time teaching it our patterns,

(03:47:55): since this app has been worked on for a couple of years now.

(03:48:00): So obviously there are hundreds of files with specific patterns that we want to

(03:48:03): keep following,

(03:48:04): and we want Cloud to keep following that it might not have by default.

(03:48:09): So yeah, I can go into a bit more of how we do that.

(03:48:13): Please.

(03:48:15): yes um so yeah i think this is the main the main takeaway for us has been that you

(03:48:21): know by coding production isn't about letting ai do whatever it wants it's about

(03:48:24): teaching ai or patterns and and letting it implement them and sometimes letting it

(03:48:28): surprise you like it did with the animation but overall if you want to ship things

(03:48:32): quickly but also consistently and at a higher quality level um you need to put in a

(03:48:36): bit of work um i think i would think of it as like

(03:48:39): onboarding someone new to your team.

(03:48:41): You don't just give them a ticket and tell them to go figure it out.

(03:48:45): You actually have to spend some time explaining how your code base works,

(03:48:50): how you decided on certain things,

(03:48:51): how the architecture works,

(03:48:53): also review their code and give feedback.

(03:48:55): So basically we've been treating Claude as an intern that just learns very quickly.

(03:49:01): And once it knows these things, it can implement things faster than any of us could.

(03:49:06): while we get to focus on the trickier integration problems again between iOS and

(03:49:09): Unity and other problems.

(03:49:14): So yeah, I think

(03:49:18): Most of the times when I see a Vibe coding or AI coding demo,

(03:49:23): I see people build things from scratch.

(03:49:26): Basically, there's an empty project.

(03:49:28): There's no constraints.

(03:49:29): Cloud can do whatever it likes.

(03:49:31): But that's not how real work looks like,

(03:49:33): especially if you're working on an app that was not initially Vibe coded that has

(03:49:37): existed for a few years.

(03:49:39): So we have established patterns and conventions.

(03:49:42): We have our own flavor of Swift.

(03:49:44): For those who are not iOS engineers, Swift is a language that you use to develop iOS apps.

(03:49:49): We use the specific framework for dependency injection,

(03:49:54): which has certain patterns that Claude needs to understand.

(03:49:57): We also have, we use GraphQL and Apollo for our backend.

(03:50:01): And basically there are, you know, years of decisions condensed into these files that are not

(03:50:07): necessarily clear even to an LLM at a first glance.

(03:50:10): So you need to spend some time teaching it that.

(03:50:13): And when Claude, Claude is really good at writing generic Swift, right?

(03:50:16): Like based on examples it has seen elsewhere and it's technically correct and it

(03:50:21): might work,

(03:50:22): but it doesn't fit into what we have already created.

(03:50:25): So it's like,

(03:50:26): yeah,

(03:50:26): it's like basically hiring someone who learned Swift and has never worked on like a

(03:50:30): real team.

(03:50:32): So we need, yeah, you need to integrate this new teammate into,

(03:50:37): your team and processes.

(03:50:40): So I don't know how much of this is visible.

(03:50:42): I will dive more into each of these.

(03:50:44): But basically, we have three main pieces in our Cloud setup.

(03:50:49): So we have agents, and these are specialized AI workers that handle complex tasks.

(03:50:55): Then we have hooks, and these are scripts that run before or after Cloud uses tools.

(03:50:59): And they basically help enforce guardrails,

(03:51:02): so blocking patterns and directing tasks to the right agents.

(03:51:07): And then we have skills,

(03:51:09): and skills are markdown files written in plain English,

(03:51:13): basically,

(03:51:14): that teach cloud domain-specific knowledge.

(03:51:16): And we have 13 of them at the moment, and we're adding more.

(03:51:21): And then we also, I think I talked a bit about MCP integrations.

(03:51:26): I call them magic cloud powers.

(03:51:28): So we integrate with other tools like GitHub, Linear, Sentry, Figma.

(03:51:33): So basically,

(03:51:33): cloud can now read a bug ticket,

(03:51:35): cross-reference it with crash reports,

(03:51:38): and then suggest a fix all in one conversation because it has access to all these

(03:51:41): tools.

(03:51:43): So going a bit more into agents that we use.

(03:51:46): So I think a tricky thing with iOS development is that since it's not in a browser,

(03:51:51): people assume LLMs actually can't build the code and see what it looks like.

(03:51:57): But this is not true.

(03:51:58): You can actually create an agent,

(03:52:00): for example,

(03:52:00): that knows how to build your app and actually gets a log of all the errors.

(03:52:06): that come out of that.

(03:52:07): And then it can,

(03:52:08): during a conversation,

(03:52:09): it can continuously rebuild your app,

(03:52:11): fix errors,

(03:52:11): build it again until it gets a working solution.

(03:52:14): So we have an agent for that.

(03:52:18): Then we have a few other agents.

(03:52:20): I think my favorite ones are the reviewing ones.

(03:52:24): So we have a logic reviewing agent.

(03:52:27): and a pattern reviewing agent.

(03:52:28): And the cool thing with agents is that you can have them assigned to different models.

(03:52:31): So for example, we use Opus for logic review, and then we use Sonnet for pattern review.

(03:52:37): Some things are just more difficult to do and more suited to specific agents than other ones.

(03:52:44): And then another agent that is great is the PR shepherd agent.

(03:52:48): So basically you put up a PR and then we have a few agents running in parallel

(03:52:52): reviewing the PR and then Claude can go back and by shepherding the PR,

(03:52:57): I mean it reads those comments and addresses them and commits changes based on the

(03:53:03): feedback there.

(03:53:04): So basically it can, yeah, it can just go back and forth with the review, the review bots.

(03:53:09): And by the,

(03:53:10): yeah,

(03:53:11): by the time it's,

(03:53:11): it's finished shepherding,

(03:53:13): you're probably in a,

(03:53:14): in a pretty good state.

(03:53:15): How are you running these agents?

(03:53:17): Like you're,

(03:53:17): you're in cloud and you're just like run the PR shepherd agent or how is it,

(03:53:20): how are they activated?

(03:53:21): Are they, are they.

(03:53:22): Yeah.

(03:53:22): I mean, they can run both locally, but also remotely.

(03:53:25): We have some machines that we use for CI and we have cloud run on those.

(03:53:30): Okay, cool.

(03:53:31): Yeah.

(03:53:33): How much are you handwriting code these days versus agent?

(03:53:37): I would say when I'm handwriting code is mostly in response to something wrong that Claude did.

(03:53:43): I usually assume that Claude will, or I usually use Claude as the

(03:53:48): as the first thing I do is I go and tell Claude,

(03:53:52): I want to do a specific thing,

(03:53:53): you know,

(03:53:53): of course make a plan.

(03:53:54): And then I'll hand write code when something doesn't fit what I thought was supposed to be done.

(03:54:02): But even then I can oftentimes just tell Claude,

(03:54:04): hey,

(03:54:04): like this,

(03:54:05): like as if I'm reviewing his code,

(03:54:06): I can just say,

(03:54:07): this is the pattern you should have followed and provide a code sample.

(03:54:10): And then it will follow that.

(03:54:12): And are you mostly using Cloud or Codex or both or something else?

(03:54:16): Yeah, we are all mostly using Cloud.

(03:54:18): We do have Codex,

(03:54:18): for example,

(03:54:19): review PRs,

(03:54:21): and it does sometimes catch things that Cloud does not.

(03:54:23): But for actually writing code, we all tend to use Cloud, either through the CLI,

(03:54:32): I've also played around with using the desktop app and cloud code there.

(03:54:36): And sometimes I do prefer it to CLI because it's a bit more of a back and forth.

(03:54:40): So I like it, for example, if I'm like working on a plan for something.

(03:54:44): And otherwise, yes, I know we have a teammate at least that uses Cursor and Opus within Cursor.

(03:54:49): Great.

(03:54:52): And then skills.

(03:54:53): So again, I was mentioning skills are basically these like markdown files.

(03:54:57): This is how we encode tribal knowledge,

(03:55:00): like everything you would tell a hire in their first week,

(03:55:03): write it down in a skill.

(03:55:04): And we're actually onboarding someone right now and have been creating skills along

(03:55:08): with the questions that they have been asking and like problems they have been

(03:55:12): bumping into.

(03:55:14): So, yeah, actually, that I think is a really interesting one.

(03:55:18): Henry Finkelstein asked who owns governance on these agents and skills like who are

(03:55:22): who's ensuring that it's doing what's expected,

(03:55:25): who's in charge of maintaining them,

(03:55:27): all that kind of stuff.

(03:55:28): Yeah, so we are a very, very small team.

(03:55:30): So on the iOS side,

(03:55:31): there are four of us,

(03:55:33): well,

(03:55:33): five of us now,

(03:55:34): since we just onboarded someone new.

(03:55:35): And we have Dan, so Dan Fetterman, who is the, I guess, the lead of the team.

(03:55:43): He's been the one really championing this and pushing this forward.

(03:55:45): And I would say he is the owner for this.

(03:55:47): I think these are things that are much easier to do in a smaller team versus a bigger one.

(03:55:51): I know that the backend folks,

(03:55:54): for example,

(03:55:54): have their own setup,

(03:55:56): but again,

(03:55:56): Dan has been the one really pushing them to use these things in their own repos.

(03:56:02): So yeah,

(03:56:02): I would say it's one person right now,

(03:56:05): and I think it's much easier to do when you're in a smaller team.

(03:56:08): Yeah,

(03:56:08): in case anyone's wondering,

(03:56:09): there's just a cabal of pro AI dance that are distributed through throughout

(03:56:13): startup land right now,

(03:56:14): just like just injecting Claude into every organization.

(03:56:17): Yeah, keep going.

(03:56:22): But yeah, but I think it's great to have someone like this.

(03:56:25): And if you don't have a Dan, you should get one.

(03:56:28): Get a Dan.

(03:56:29): Yes.

(03:56:32): And then I mentioned also hooks.

(03:56:34): So hooks are more of these automated guardrails and they run before or after Claude uses tools.

(03:56:41): And basically they make sure that we're not doing things that we have noticed it do

(03:56:47): before that we don't necessarily want it to do.

(03:56:50): So for example,

(03:56:51): I think these are some very specific to our work stream,

(03:56:55): but yeah,

(03:56:56): they've come up as we've been building.

(03:56:59): We just basically want to make sure that Claude can't accidentally do the wrong

(03:57:03): thing,

(03:57:04): which you sometimes have to do with these tools.

(03:57:07): And then,

(03:57:08): yeah,

(03:57:08): I mentioned earlier,

(03:57:09): we use MCPs,

(03:57:11): model context protocols or magic cloud powers,

(03:57:14): and this gives cloud access to our external services.

(03:57:16): Again,

(03:57:17): I would assume these are things that are much easier to do in a smaller company

(03:57:20): where you don't have to go through weeks of approval to get access to do something

(03:57:25): or approval for an integration.

(03:57:27): But yeah,

(03:57:27): we basically have,

(03:57:28): so GitHub is the essential one for reading and creating PRs and then reviewing them

(03:57:32): and shepherding them.

(03:57:34): Linear is great.

(03:57:35): You can tag Claude in a linear ticket and it will open a PR solving the bug that came up.

(03:57:42): We also use it for Sentry, for crash logs, post hog.

(03:57:47): We can use it for analytics.

(03:57:48): We have a channel where we just tag Claude and ask it to analyze something specific

(03:57:53): and it will usually create a graph that's pretty solid.

(03:57:57): And then, what else do we use it for?

(03:57:58): I guess for Datadog, for monitoring.

(03:58:00): But basically,

(03:58:01): yeah,

(03:58:01): I think these really enable cloud to work across systems,

(03:58:04): which is how you as an engineer work.

(03:58:07): So, and you can access all of this in one conversation.

(03:58:10): Like you can read a linear ticket,

(03:58:11): you can pull the stack trace,

(03:58:13): then check the Figma component,

(03:58:14): for example,

(03:58:15): is it just a fix,

(03:58:16): create a PR,

(03:58:17): and you don't have to like context switch yourself between all these apps.

(03:58:22): Amazing.

(03:58:23): We've got one minute left.

(03:58:25): What do you want to leave folks with?

(03:58:29): Yes, I guess this is the main thing.

(03:58:32): Ask Claude to help write your cloud setup.

(03:58:34): It is actually really good at that.

(03:58:35): You don't have to handwrite it yourself.

(03:58:37): If you'd explain it to a new hire, make it a skill.

(03:58:40): Different tasks need different types of specialists.

(03:58:43): You can use Sonnet for some things, Opus for other things.

(03:58:45): Who knows what models are coming out next.

(03:58:47): And MCPs are your MVPs and help you work across streams.

(03:58:52): And I'll just jump.

(03:58:54): Yeah, I've talked about designers opening PRs.

(03:58:56): This is great if you can do the Slack integration.

(03:58:58): Again, bi-coding can be used in production.

(03:59:02): And, you know, your non-engineers can ship code and your users won't know the difference.

(03:59:08): And that's it.

(03:59:09): I just wanted to do a quick pitch that we are hiring at Portola always.

(03:59:13): If anyone's interested in checking out the roles we have available, I put up a link here.

(03:59:19): And yeah, if you have any questions, you can find me on Twitter at Paola Rambles.

(03:59:23): Also reach me via email, paolasportola.ai.

(03:59:27): And yeah, this is wonderful.

(03:59:28): Thank you so much for having me.

(03:59:30): Amazing.

(03:59:31): Thank you.

(03:59:31): This is amazing.

(03:59:33): I learned so much.

(03:59:34): Everyone go check out Paola on X, check out Portola.

(03:59:39): Yeah, thanks a lot.

(03:59:40): Thank you.

(03:59:42): All right.

(03:59:43): So now I am very excited to add Brandon Gell, the legend.

(03:59:50): How are you doing, man?

(03:59:51): Good.

(03:59:51): I've been psyched watching what's happening here.

(03:59:54): It's been really fun.

(03:59:56): So for people who don't know you,

(03:59:57): you are the COO at every you are like,

(04:00:01): well,

(04:00:01): this would none of this would happen without you.

(04:00:03): You're so important to everything that we do.

(04:00:05): And just yeah, awesome.

(04:00:07): So where are you?

(04:00:09): Where are you?

(04:00:09): Where are you calling in?

(04:00:11): I'm in the Black Mountains right now in North Carolina,

(04:00:15): which is great for focusing because there's nothing around me.

(04:00:18): I went for a hike this morning.

(04:00:20): It was a 30-minute drive to the closest hike, which makes no sense because I'm in the mountains.

(04:00:25): But yeah, that's what's up with me.

(04:00:28): um and we got uh we got alex and cj from 10x joining right yes i don't know about

(04:00:35): alex um okay but i am going to i'm going to pass it off to you but before i do that

(04:00:41): if you are here uh you should know every is the only subscription you need to stay

(04:00:46): at the edge of ai

(04:00:47): If you subscribe, you get access to ideas, apps, and training.

(04:00:50): On the ideas side, we have a daily newsletter.

(04:00:54): Whenever anything new comes out in AI,

(04:00:55): we have a day of vibe check where we do hands-on testing to tell you what you need

(04:01:00): to pay attention to and what you don't,

(04:01:01): and to teach you about how we're using AI in our work and in our lives.

(04:01:04): We also give you access to a suite of apps that we've built for ourselves that do

(04:01:11): everything from do your email for you to organize your files for you to do smart

(04:01:16): dictation for you to be your AI ghostwriter.

(04:01:19): We use it ourselves.

(04:01:20): We build it ourselves and we give you access to them as well.

(04:01:23): All part of the subscription.

(04:01:24): We also do these trainings.

(04:01:26): So we have live streams like this one.

(04:01:28): We do twice a month camps where we teach you about how we do our own workflows inside of Every.

(04:01:33): And we also bring in people from top companies like OpenAI and Cursor to teach you

(04:01:38): their own power workflows as well.

(04:01:40): That's all available for one subscription.

(04:01:42): Brandon, is there anything I'm missing?

(04:01:44): Any other reasons someone should subscribe to Every?

(04:01:48): I just think if you're excited about what's happening right now in the world of AI,

(04:01:51): there's no better place to dig your teeth into.

(04:01:54): There's so much happening here every day.

(04:01:56): And we build incredible products that you should probably use.

(04:01:59): Absolutely.

(04:02:00): So I'm going to add CJ to the stage, and I'm going to allow you two to have a conversation.

(04:02:07): All right.

(04:02:07): See you in a bit, Dan.

(04:02:08): See you later.