A (7:29)

So what you're saying is the alternative is you have to be ralph, right? You have to be that person to be like, yes, no, yeah, you have.

B (7:37)

To do all the testing, right. And that's a pain. So. And the agents are smart enough now, Opus 4. 5 is smart enough now that if you give it good acceptance criteria and it can do it. So, all right, so we've got a prd, a product requirement doc. We've used this skill. Let me take this here. Okay, so now this is called the RALPH PRD converter. And it says, it converts a PRD to a PRD JSON file for the RALPH autonomous agent system. And it basically says, okay, take a PRD and markdown format and then output it like this. Right? So this is telling the agent exactly what you need the PRD to be turned into. And it has a couple really important things here, like the story size, the number one rule. Each story must be completable in one RALPH iteration. So this is the other big unlock with working with agents is that they have a context limit, right? So you know, with Opus, you're looking at about 168,000 tokens. You have to be picking chunks of work that can be fully completed within that context window. And this is why RALPH works so well, is that it runs little independent threads where it completes one user story because it's not very big. So you've got a bunch of very small user stories. Story ordering, it puts the ones that it should do first at the top and then it says acceptance criteria. They must be verifiable. Right? So this is really key. So good criterias are add, you know, status columns to the task table with default pending. You can filter dropdown has options all active completed. And again, these are things that amp wrote for me and I verified, but they're important that it's in there. So let's go back. All right, so we've created our product requirement doc. Great. We've used the RALPH skill to turn it into a JSON file. Great. Now we run a script. So this is a little different. So most of us who have used agents like an AMP or a cloud code or a cursor, you're used to opening up that program, right? So you open up cursor or you open up AMP and you start talking to it, right? This is actually a BASH script, right? So this is a file that, that your actual local computer runs. So I'm going to show you actually what this looks like really quick.

A (10:19)

And for people who don't know what BASH is, it's. It's just a way for.

B (10:24)

It's just a type of. So. So, so when you open up your command, your terminal, which is basically a text input box for your computer, you can type commands on it and you can also run scripts on it. So that's what a bash script is it's a type of file that the computer can run from the command line. So I'll show you what this looks like. And again, this is open source. And we're going to have all these notes, all these links in the notes. And so you could literally take this RALPH public repo I have, download it and use it. So let's go and look what this actually looks like. Now, this looks kind of scary, but let's look at what it's actually doing. So it's saying, okay, how many times can this script run? The default is 10. It has a script directory, it has the PRD file, right? Like we talked about this. So it's called PRD JSON. It's got a progress file, which we're going to talk about. And then the idea is that it archives the script when it's done, and then basically it does a loop. And I'll explain how that loop works visually, because it looks nicer than Bash, but this is the actual file that runs. So once again, we'll quickly recap. We've got a PRD rewrote. We converted it into a PRD JSON, which is just a list of user stories. Every user story is very small, right? It has very clear acceptance criteria. And then you go to the command line and you run the RALPH script, which is you type a couple characters and you hit enter. Now then what happens? And why is this so exciting? So, all right, so then what happens is the agent, you know, I use amp, it picks one of the stories, usually it picks the first one and it's looking for any user story that has passes false, right? So you can see here, has this user story been, does it pass or not? No, this one doesn't. So let's grab it, right? And this is the other really cool thing about ralph. So for decades, teams of engineers have been working this way. You have a list of usually sticky notes, or now that's Kanban boards, and you have a user story. You pull it off and you say, I'm going to do that. You go to your desk, you start coding, and then you complete it. You commit it and you merge it. Then you come back to the board and you grab another sticky note. This is the way humans have been coding forever. And the reason why is because it works. You have a unit of work that you can understand, that you can test and you can complete independently, right? And this is exactly what RALPH is doing. So it's picking a story, grabbing off the board, and it's tackling it. So then it starts working, right? This is what we're all used to. If you're in cloud code, or if you're in amp, or if you're in cursor, you say, okay, build this thing and you start seeing it working, right? So I'm actually going to show you today I posted and I said, okay, this is a real thing. I actually built an entire feature with ralph and these are the steps I took. I created the prd, I created the user stories, and then I started ralph. And then I'm just going to show you an example of kind of what this looks like for real. So this is a thread and amp, this is basically what you would usually look at and type into. And this is the system prompt that RALPH gets, right? So it says you're an autonomous coding agent working on my project, your task, read the prd, read the progress log and just do a bunch of stuff. And then it basically says, and then it starts working on it. Do, do, do, pretty straightforward. And then you can see down here it's actually cranking, right? So this is the kind of stuff you're used to seeing. So here it starts writing code. Now the whole time that this is happening, I'm sleeping, right? Or you know, I'm having dinner with my amazing wife or something, or I'm paying attention to my kids, right? But RALPH amp is cranking on that first user story without me having to look at it, give it feedback, do anything, because it has everything it needs the context, it has the acceptance criteria, it has everything.

A (14:49)

What about tokens? Is it burning through my tokens?

B (14:53)

No, this is the beautiful thing. So I actually looked at a couple of these. Let's actually look really quick. Yeah, so look at this. This was three bucks. I mean, $3, that's less than a latte.

A (15:07)

Totally.

B (15:08)

So the question is, how valuable is your time, Right? I think that, you know, for some people three hours is going to be too much, right? So for instance, the typical RALPH cycle is probably 10 iterations, so you're looking at maybe 30 bucks. But think about how expensive and how hard that would have been with your time or developers time. So it becomes pretty amazing. And I'll do one plug, which is that we're about to launch a daily free token allowance on AMP. And so you get like 10 bucks a day. So you could do this for free. So, and so I think people are afraid of like, what if the agent runs by itself? Is it going to go off and do crazy things? And the answer is no, because you gave it a clear User story with clear acceptance criteria. Right. It's going to be actually a pretty small thread, so.

A (16:01)

Right. I mean, it only does a crazy thing if your criteria is crazy, right? Yeah.

B (16:07)

So this is the whole point is you have these very small atomic user stories. So. All right, so Amp cranks, it actually finishes the feature, the user story, and then it commits the change. Right? So it's like, okay, I did everything. And this is really important. This is in the system instructions for RALPH when you finish commit, because you may need to roll back, right? You may need to go back and fix things. And this allows the agent to do that if it needs it. This is pretty standard stuff for agents. And then what it does next is it updates the PRD JSON. And if you look at this, it's going to change this from passes false to passes true. So all of a sudden it's like that person, they go back to the board and they put the sticky note back on and it's crossed off. Right. And then this is one of the most important things is that it logs its progress. So this here is really important. So what's going to happen is when this iteration is finished, it's going to learn things. It's going to learn things about your code base. It's going to run into walls and have to go around them. You don't want it learning that stuff every time. Right. So what it does is a very simple prompt. It says if you learned, if any of the files you edited have Agents MD files in that folder and you learn something that's important, update that file. So not only are you, and this is what Kieran and the team over at every absolutely crushed with their compound engineering concept, is that your agent should be getting smarter every time it makes a mistake. And by updating Agents md, you're gonna get that long term benefit. This isn't just during this iteration. You're gonna benefit every time from now on that you use AMP or cloud code. So.

A (17:56)

And for people who don't know what Agents MD is and the importance of it, can you give a quick primer on it?

B (18:02)

You bet. So what? It's a very simple file. It's literally a markdown file. And think of it as notes that you would give to a new developer who had never seen your code. And the neat thing is you can have an Agents MD file in every folder in your entire repo if you want. And what happens is AMP will if AMP starts looking at a file in a folder and it sees an Agents MD file in that folder it will read it first. So it's cool. It's like sticky notes everywhere, right? So it's like if you have some knowledge about this part of the code base and you really want the dev to know, hey, before you mess with stuff, read this. That's what Agents MD does. So you can have it at the top of your folder directory so that gets read every time the agent works. Or you can put it, bury it down in some subfolder so that it just reads that when it's editing that file. So that's really, really important. If you don't do this, you're going to have a bad time over and over again. And then what it does is it updates a progress TXT file. This is like short term memory. So this is like saying, hey, during, during these 10 iterations that Ralph is going to do, what are some things that we're kind of learning? What are we doing? How does all this work? I'm actually going to show you. Actually, I won't show you an example because it's a bunch of text. But the idea is for each iteration it's going to say, okay, here was the amp thread that we used, which is great because then the agent can go back and read it. So say that you get to iteration two and it wants to see, well, what I do in iteration one. It can actually read that thread and figure it out. It also puts a couple notes in there like, oh yeah, I probably shouldn't do that in the future. Or I should think about this. It's more of a short term memory, whereas Agents MD is its long term memory, things that it really needs to remember. So that is logging in progress Txt anything confusing so far?

A (20:05)

Crystal clear.

B (20:06)

Nice.

A (20:07)

Yeah.

B (20:07)

Okay, so what happens next? Well then RALPH says, okay, is there any more stories? So it goes back up to the little board and looks for more sticky notes, right? And it says, oh, yep, there is more stories. So I just do it again, right? I pick a story, it usually picks the next one in the JSON file. But the nice thing about it is it's like a human. It can also figure out, well, I could do that story or that story. I'll pick this one because that would be smarter. So it grabs it, then it implements, commits a change, it updates the PRD JSON file, it logs its progress updating any Agents md and it just does this over and over again until da da, it's done. So this I'm going to show you again. This was a real example of me using RALPH today. So this morning I built a big feature into my app. So I created the PRD and then I created the user stories and then we cranked and we started doing this. Now this Ralph iteration took 14 iterations to do just to kind of zoom into like one of these, right? What does this actually look like? So again, same system prompt. Now this is the key is I think people don't get it. Like, well, why is this different than me just like using AMP or CLAUDE code like on my computer. And what's happening is you're getting a fresh loop every time. So you're getting a brand new thread or a brand new instance of CLAUDE code every time. So you're starting fresh with a brand new context window. Starting from fresh. So these are the instructions, this is what we're doing. Read the prd, do the stuff, progress report format, append to it. Here's how to append. So the things that's putting the progress file are the AMP thread so that it can read it again. What was implemented, the files that were changed, any learnings for future iterations, you know, patterns discovered, gotchas, useful context and then it kind of cranks through and then it starts down here, you know, and that's pretty much it. I mean, it's very, very simple. And with that you can ship massive features while you sleep. And y', all, I've been, you know, I got my computer science degree 26 years ago, right. I've taught a million people how to code. Right. I've started companies with huge engineering teams. Right. This loop is basically an entire engineering team. Why you sleep. It's unbelievable. And this just wasn't possible before Opus 4.5. I think with Opus 4.5, this is absolutely the real deal. And just wait till Opus 5 comes out. I mean, it's so, so exciting to be alive right now.

A (23:07)

Yeah. And I think it's not just an engineering team, it's a high quality engineering team that, you know, runs by best practices.

B (23:16)

Right. For $30.

A (23:18)

For $30.

B (23:22)

It's so crazy. Now in a perfect. The real world is. So let's look back at this thread. At the bottom, I said after I finished, I tested manually and did find a few edge bugs. This is normal, right. You're going to probably run into a couple things that don't quite work right. Or aren't perfect. But AMP and I just worked on it and cranked out fixes really fast and the feature shipped this morning. So that's what all the hubbub is about with ralph again, it's Exciting because it's so simple. It's a loop, right? It's grabbing a user story and doing it. It's having clear acceptance criteria and then it's writing down what it learns so that it doesn't make the same mistakes. The last thing I will say very, very, very important I'm going to zoom in is these two steps. Writing a PRD and converting them to user stories. This is where you should spend a huge amount of time. Like you should spend an hour on this, right? It's very, very, very important that you get your PRD right and that your user stories are small atomic and that you have clear acceptance criteria. Because if you don't take your time on that, you're going to get 10 iterations of Ralph and you're going to end up with something that's not very good. And I will say one tip, one really good at the coal face with your pick as you're working is figure out how to connect your agent to a browser. I have a specific skill which is in. Let me actually show you GitHub. Let's go to Snark tank and I'm going to show you this repo called AMP skills. If the you can download all these, these are open source now not open source but this one is what I would use which is your web is your dev browser. Okay. So dev browser used this because what this does is this allows AMP or claude code to actually use your browser and test and your user stories that involve front end code remember that the agent needs to be able to test that. And testing browser is hard for agents. So you have to give a specific skill like DevBrowser. This is a free skill that a friend of mine made. It's really good. So one small tip for you.

A (26:03)

Cool. We'll include that in the show notes. If people want to get started today, how do they get started with ralph?

B (26:11)

So I would probably go to this repo. So just go to GitHub.com nartank Ralph, it's completely free. Look at the code but just tell your agent to look at this code. So tell your agent to go set it up if you're using amp. Great. And it will crank it out, get it all set up and then say okay, I want to ship my first feature. How do I do it? And it will walk you through it.

A (26:39)

And you believe that even if you're non technical you can do it.

B (26:44)

Yeah, I think you need to be curious. I think you need to have agency. But I think if you have those two things which you probably do if you're watching the show. You can do this now. It helps to be technical, right? There's a couple things that are useful, but I think we're now in a world where if you want those skills, you can acquire them very, very quickly. And I'd encourage you go out there and get your hands dirty. Like, that's the most important thing is use the thing, figure out how it works, and then if you don't get it, ask amp. Ask cloud code. Like, you have this superhuman tutor to help you. So just ask them.

A (27:23)

Ryan, thank you so much for coming on.

B (27:25)

You.

A (27:26)

You explained this exactly how I wanted you to, crystal clear. Unbelievable. I would love to have Ryan back on the pod if he. If he would come back. But, you know, would you a. Would you come back on the pod?

B (27:39)

Let's do it. I'd love to. Let's share as much knowledge as we can.

A (27:43)

And I want that comment section on YouTube Buzzing for Ryan for. Because, listen, it might look easy now. You might be like, oh, I get Ralph. That was super easy. But it's because he explained it in a simple way, which I think is the beauty about you, Ryan. So thank you so much. I'll include links for all of this to get started. Links to follow Ryan on X, you know, links to amp, links to Claude code, and Ryan, anything else you want to make mention to.

B (28:13)

Well, thanks, Craig, for your content. Like, it is a big unlock. I think if people want to learn these things, they can. And your content is a big piece of that. And the second thing I would say is you do not need a computer science degree, y'. All. You can do these things. Like, if you are curious and hardworking, you can now do anything. And now's your moment. So if you've got an idea, build, build, build.

A (28:38)

Yeah. Quit watching us right now.

B (28:39)

Get out of here. Go. Come. Get out of here.

A (28:42)

All right, we'll see you next time. Thanks again, Ryan.