Notes provided by Gemini. We were all using the same microphone, so many people were speaking, though J. Langley is listed as the speaker.
J. Langley: the environment under familiar withd. You found it. Sign in, please. Oh, yeah. Yeah, we’re at a real like facility that we have sign in and stuff. So, Unity 3D has a UI toolkit. It’s called that’s one reason I chose this is the UI toolkit language is almost exactly CSS. So AI is very good at writing CSS. So it’s like hey why don’t I use that as a framework? Most people use that for UI development but entire dig. So first experiment can I just straight up prompt code? Second experiment is I don’t know how much you guys are familiar with the game industry but there’s a lot of hate for AI created content in games.
J. Langley: So my other goal is for the game I make to not flag anybody’s AI detector. I set a goal by October. That’s when Steve has one next. I want to have a demo out on next October. So you guys are catching this in the first I have about 20 hours this so far just doing a little bit of prompting each night after this will be steamless. So I wanted to avoid a lot of standard AI for art assets. People are very good at you can do that if you want what is AI generated art assets for games they’re not so good at finding out that code is written by so I chose something that’s entirely unless it doesn’t work people just think the game sucks it doesn’t matter can you guys see the screen kind of get dark ambience. It’s how developers like it anyway. We need more dark. That’s that’s acceptable. And it really throws me off because you’ll see in reviews people are actually complimenting games that use AI.
J. Langley: They’re like, “Oh, these graphics look incredible. It’s a small dev team.” Oh, I found out they used AI and it sucks. Don’t lie. Why are you complimenting? Anyway, that’s that’s another uh So, I won’t show you the original problem. Kicked everything off. there’s no power or anything. I do have a background in game design. So my first step was just working with quadopus 46 and I laid out you know every step of the way to design a game was create a base plan for it to work off of. I needed to find a game that could work off of UI. So what I ended up picking was I had a friend that just come back from Disney World like that talking about the Star Wars exhibit spaceship. He said they always pick kids for like the captain’s seat and then an adult get stuck in the engineer. Was that that’s me? Yeah. And all the engineer has is one button to press. And I’m like that’s not fun at all.
J. Langley: Engineers just have kind of a cool job at sci-fi spaceships. So my idea is you serve on a mercenary spaceship. You don’t get any decisions like you’re just the engineer ship in that job. Something else you’re the spaceship. You don’t decide where to go, what to do. You just have to figure out how to keep power going to all the systems and keep it supplied through Wi-Fi pretty much. Yeah, you’re Scotty from Star Trek. So I laid all that out for clawed and basically just left it up to it to kind of interpret how all of that so I can give feedback and because it’s not going to be that long talk what I’ll go through and say is it’s been able to do almost everything on its own. uh there’s very few connectors that’s happening all week through and literally cloud code just spelled out correctly how to do that. So for example, if I have a manager set up, I might have to drag in UI document create and it’ll just say hey do this and I’ll have two seconds of work and I’ll keep running next.
J. Langley: Every single bit of the structure here was created by FL. I’m doing a standard kind of rogue light setup where you can get you know different kind of weapons and engines for your spaceship is the ambience. So it has successfully come up with all the different variables for it. And it’s created all of these different randomized items for me to it. Something that really surprised me is when it came to the audio created a tool to generate see if I can entirely thinking I’d be showing this off this early. Uh, but right now I just have placeholder audio, but I didn’t even think about claw being able to do that. But it can literally just write the wave files. So it wrote something in Unity to generate all the wave files for my placehold and I’ve been able to edit itself where first it just had to generate, you know, real simple beeps and loops. But uh let’s see if there’s I don’t know if there is there any audio in here. Yeah, I doubt it.
J. Langley: Or maybe there is. So that’s a little sound that it made for venting the reactor to make like a hissing reactor sound. Really good at little flicks. Very old school, but I’m still impressed. Uh, one of my next steps is to feed it uh more instructions. So I found if I tell it to actually read uh first principles design notes on audio that all of a sudden it’ll start refining better wave files for me. So again I know I could just feed it in I could spend $5 buying a bunch of assets but really I’m trying to get everything out of here. So what impresses me, this isn’t perfect, but after about five hours, uh, I had a working UI start. And one thing I requested was kind of an old CRT. This probably won’t show up while we’re here, but see like I want a little barrel view for it. Scan lines, color shift, staticky, noise. There’s an overall intensity for everything setting. And a lot of these I’m going to end up using for like your ship takes a hit, so everything will start flashing and shaking around.
J. Langley: But all of these shader effects were written by flaw code. All of this was tied into it. Same thing with the audio settings. All the audio settings work UI for the game. You didn’t really describe settings, right? Just came up with it. No, I told it what I wanted, but that’s it. So the first thing I said was I want to emulate like and they did okay all of these except for the last three wave ones and then I requested to add those in kind of see the edges. That’s cool. Is pod able to see these results? Like does it screen capture or anything just working correctly or just uh starting off I just did complete touchsc I have done some screenshots just you know I really am trying like barebones as I can and then just stepping up a little bit each so I’m currently sharing uh but But you can tell it like oh I don’t like how like these buttons overlap or whatever right and then it will go and say oh yeah realize it overlap and sorry and fix it right yeah it biggest problem was UI toolkit is so close to CSS that started just right CSS all itself so this is it keeping track of what are the steps But there’s also instructions in here about stop using CSS and it’s following it pretty well.
J. Langley: Not having to do any skill so far. It’s it’s learned now. Hey, let me look at the official documentation. And within, this is a little advanced past it. Within about five hours of work, I had what could technically be called a game, even though it’s not whatever want to play it, but every single element, everything. And you’ll see like a little graphic placeholders for shifts. Um, save system works. Crashes. You can pick up. Now, the idea was you would just have your engineer panel uh with all the different systems you were controlling. There’d be a main reactor that tracked your board usage. Down here you see a cruise comp with all these messages coming in because it ship showed up and right now buttons overlap. So the last change I put through power all of these different systems and you’ll see up here is the enemy ship. I don’t like displacement currently, but you’ll see its health start going down as our weapons hit it.
J. Langley: The more the weapons are charged, the faster we shoot it, the higher our shields are, the less damage we take from it. Every system is wired up working correctly. If I left life support off, the crew is well, that’s it. That’s not engineering. All of these little placeholder graphics, the little sine wave reactor these were all the start placeholders I thought. So the other next test audio is I am going to go to just for an example sticking with text only I took a screenshot and just dropped it into Gemini and asked it to be pretty nice. So what I’m working on now is I’m dropping this into cloud code saying and I want to see can you can it actually use the me and everything else to recreate this. I don’t know if it can but probably close. Looks like the code could get really big. Uh so I don’t have debug turned on right now. So, it’s going to take a little second for us to kill this ship, but uh you can have a win condition, a lose condition.
J. Langley: We’ll get a little popup once the ship destroyed. There’s 15 different stages. So I mean all the same hours it’s me walking by my laptop on the counter see if my next coming back half an hour to check on it. How much have you spent like just in normal monthly? So this is off of the $20 a month. $20. So usually in about two hours each night I blow through all the usually I get off the work half an hour later my laptop. That’s that’s part of the test. So apparently three of my crew members died because I wasn’t paying attention. Kind of blown me away. If you haven’t worked with Ruby before, it can be a little bit tricky to make sure everything is correctly tied up and everything’s feeding into each other. And I have had zero zero input required on here. The organization actually follows really principles. So you haven’t touched any of the zero. All I’ve done is Have you looked at any of the code?
J. Langley: No. Okay. Because I’ve got half of that answer. So, I I have actually on mine that I’m going to talk about, I’ve actually gone through and looked at stuff and then tried to guide it a certain way, but it didn’t occur to me to just go with it. And I’m also so everything is locked into a single folder. I’m not even using Git or anything else to back it up. I’m using old school That’s new. Is the folder is the name of the folder yellow? Project new project new, right? New new 2026. issue any sort of testing loop at all or is it like just you ask for changes and then you just So yeah, as soon as it makes a change actually is telling me what it wants me to test. Okay, tell him that you’re not testing and he needs to do it. And then when I recognize something that I know is offer will correct problems. But the most I’ve had to do is down in the indies console.
J. Langley: There have been maybe three times I’ve had critical errors that won’t just copy paste. It’s always been a one shot fix. So again this is a very simple format. I picked it very intentionally for the project to do but so far it’s doing everything. Uh the name itself sort of like yellow truthno is I think polish and it’s basically what are you gonna do like it’s out of my hands which I think fits the game idea stuck in a situation. There’s a probably but it’s it has built into it vision model that will take in anything that you annotate. So say for instance you can screen capture also do a a record a screen recording and you can put make notes and comments or um say you know this is how it’s functioning this this is what it’s doing this is not lining up I need to fix this I need to fix that and then take that entire thing and then drop that back into block and then it will build out the correction control. So you can actually look at what you’re capturing and then it will also control computer.
J. Langley: So you can you can actually make it do the smoke test for the system. So that’s you will burn it through your token. Yes, you will. You can do anything but you will burn through your token. So something I forgot to do at the beginning. I went ahead and just put through next prompt. I was going to prompt it should do sleep. So figure by the end of the night on your nano banana. I’m assuming that you use Gemini for the signing. How much prompting did you do to get to that final sign? Oh, so I put in the screen of it was currently better UI for spaceship engineering games while keeping the same basic elements and layout. Follow first principles. Make each individual system box better reflect the system controls. keeping it now. Yeah. Was first principles a key phrase for something you’ve loaded in through an agency or something or was that just all you had to say? I’ve gotten really good results out of first principles lately if I’m using the latest uh 47 version for some reason that raised by itself.
J. Langley: Yeah. I was curious if you’re triggering something else and it’s ancient city or something. These are the first principles. Yeah. No, it’s uh it’s just something I found recently that telling a generally steers away from the genericator like, “Oh, here’s here’s something that might work.” So, a little more effort. Yeah, try please. So, yeah, that’s Any other questions about the silver ball project? Hopefully uh October this year I can do something saying, “Hey, this is out on Steam. Here’s what it looks like.” Then you guys will hopefully be the only ones to know that it’s 100% AI. If you want to introduce some other pages or you have any idea what other type of pages to the project uh like to introduce web cloud code or just other AI in general what about cloud code like say you can’t this ticket manager create this ticket then there’s like a manager honestly the the biggest quality of life I’ve thought about is I mentioned just like a standard old tool like saving all that I have tracking updates and all that.
J. Langley: I thought about just putting together a little SQL bite for sure have it automatically track changes. I thought about doing that just quality of life. I don’t think about just code. I haven’t I haven’t really let myself think about other agents or get too serious. I kind of do that at work. So this is my like you know what what can I do five minutes at a time? What’s the most annoying thing that you’ve run into with this or the thing that did something that you didn’t expect it either you expected it to work and it didn’t or you expected just you know what I mean? I don’t know if I’m asking that right. Yeah. So the current graphics problem I have been trying to fix. I think a lot of you know how flazy sometimes like a lazy attitude on things. It has basically given up on how to properly size different areas. That’s a pre-existing. Yeah. So when I mention like this little field, you know, the little display they orient, you know, notice how that’s here.
J. Langley: The little pips on the left are here. Uh this little layout it picked at the very beginning and it does not want to break out. I cannot have it just like make something wider without reinforcing it. So that’s what my current drive is for is like recently it said, “Oh yeah, a way I can make that display window. I’ll just make it taller.” So you’ll have to redo the whole thing. You know, I’ll just make more room for it. Let’s just make more room for this element. It’s not fixing a really good uh solution for quad or codeex or anything lazy is a it doesn’t involve actually reading the code unfortunately. Uh, but it basically I do all of my stuff through merge requests because I have a GitLab instance at home and I basically say I’m going to read through it and if this is not here, I’m not going to merge it. And so everything you just did is not it doesn’t matter. So by putting that gate of like it doesn’t matter, you know, you can’t drink meat, you know, so it’s not going to get done.
J. Langley: And that kind of forces it to keep going until it has that because it knows it’s not going to hurt and it knows that it cares about that carrot. Yeah. You know, it’s I’m just I’m throw away. I’m trying to be helpful. The model screenshotting it and then getting back to it and it’s still picked. Yeah. So, that’s what I’ve just gotten into like about the last week or before that I was just describing. Oh, yeah. You’re right. Yeah. As you continue to build, how much context are you putting back in to do prompts? Are you trying to conserve tokens to stay in the $20 limit? Um, does does it keep growing and you keep having to give it more or does it like you’re going to adjust the page? Does it even need to know what the original bubble was? So, you’ll see in here the only thing in this interface, these are all different chats I’ve been through so far on it. And usually I just go until it compresses and I don’t think about it too much.
J. Langley: It compresses I get a hand off comp. I’m just like hey after everything I already have it locking file I just have it generate and usually the handoff is not read this file step. So you have told it to like keep running log of what’s in so that you can feed that back. Yeah. And it’s been that’s what all these little docks are. Can you just click on one of them? Yeah. Yeah. This stupid human luckily never reads this crap. So this is this is one of the few files. This is one that uh cla it used to be called all system screen. Um this is what it created when I described uh overall game design. She did. This is one of the first ones. Yeah, this is the first one. The very first. And then for later one like here’s read carefully. Yes. Well, the hand of step 33 probably had a lot of this content already.
J. Langley: Okay. So, this gets getting a little bit bigger perhaps. Okay. Okay. Thank you. Oh, I do it. whether that’s really helpful or not. How often do you stop it or redirect? Never. Okay. Literally, I’m not even looking at this much. Like I said, I put in the prompts and then half an hour later come back. There’s a button I need to click. Yeah. Click yes. Click yes. Let’s not I’m using at this point. Okay. Yeah. Sure. Move all my stuff over. Yeah. Wow. Yep. See if this All right. There we go. See if this is good enough to drive. All right. Hopefully this works. I’ve never actually There we go. Never actually driven with this laptop in this room. See, don’t forget how to do everything as soon as you start to present because that’s how life works.
J. Langley: Our screen. Yes. All right. Thank you guys for listening to my very Okay. So, I am presenting. So, um in a similar context, I don’t know if I can remove this or not or at least get it out of the way. So, uh no dark mode. Sorry. Um I guess I could ask it for a dark mode and it’ll just do it. Um, so starting at the beginning, uh, I found myself over Christmas driving through, uh, cotton fields, cornfields in Mississippi, uh, visiting family while listening to an AI podcast about vision and new vision models and world models and all kinds of stuff. And it occurred to me, wait a minute, um, I may be able to drive a robot visually instead of having to rely on GPS. Because if you don’t know, when you actually have the leaves of a crop connect and cover, um you’re pretty much in an RF restricted uh place underneath. So, uh precision GPS. Yeah, go for it. Fine.
J. Langley: Um that works above the canopy, but nothing. Anyway, so I started playing around. Um hopped on a uh a voice. Of course, I’m driving, so I hit Gemini, flip it to voice mode. I’m just talking to Jim and I for like three hours, building out business plans, uh, doing research on who already does this kind of thing. Well, how many acres of corn are there planted in Mississippi every year? Well, what do they spend on chemicals? Well, what do they spend? I mean, just hours and hours of that that I now have in kind of a knowledge base. Um, I have had to take that whole thing and then this is the fun thing. Take that. I threw it over to Claude and I said, “Hey, Jim and I did this and I don’t trust it.” Um, it probably has, not that Gemini matters for it, but uh, I did tell it I I am worried about confirmation bias and a model telling me what I want to hear rather than actually taking the time to evaluate options.
J. Langley: And so I got some interesting things from that and a gap analysis where it had misled certain ways and while true and um, I love Claude’s uh, uh, the the word honest shows up a lot in those responses. Well, here’s what it said. But the honest answer is this. And I’m like, correct. I appreciate that. Um, so then I started playing around with u and this will go uh public when we post it, but it doesn’t really matter because um I’ve learned that if I can build this out with clawed code and the amount of time I’ve have, so can anybody else. Um, you’re at a point where your bright idea um is executable within a month or two. Um, so the the actual keeping a lot of stuff secret doesn’t really matter as much anymore. Just go do it. Um, and if it didn’t work, throw it away and do something new. Um, so I started off um with Gemini because uh with the Huntsville AI account, we’ve got the uh Google Workplace and Gemini available for like $13 a month because we were on one of their early plans or something.
J. Langley: So I already had it. So sure, go play with that. And so I was uh I was typing in props, getting things back from that, copy paste into VS Code and try to run it to see what happened. I I quickly got off of that and started using Open Code. Uh and then Klein, which we’ve used uh a couple of times here. Uh pretty much using Open Router and pointing it at the cheapest model I could get away with that would actually do the job. And so I was working through that and I think we were at the AI symposium I think January. I think it was Leaf said, “Hey, I’ve got a seven-day trial for Clawude Code.” I’m like, “Yes, please.” And so he threw me a token or something and I set it up on a I think a Thursday and by Sunday I was just completely blown away. um and out of tokens at least twice or whatever it’s uh whatever it’s called when you run out of your plan because uh uh I I then spun over to a $20 a month plan similar what you’re you were running.
J. Langley: Um and uh what I’m some of what I’m going to talk I’ll show you what we built and kind of why and then some other fun things. Uh but kind of what I wanted to talk about was the difference between uh building a project out with Claude Code versus my day job um which is a large modeling and simulation contract for NDA. I’m typically working with 30 to 50 software developers all working between I think seven eight different agile teams all coordinated together working doing a bunch of stuff. And so it gets really really interesting when I know how to do that with real people. Um, and then how frustrating it is when I have clawed code just blazing like crazy and I write something that would take an agile team two weeks to implement and it does it in 15 minutes. What do you do? I I got to come up with another story. I got to give it more work. Give it more work. Give it more work. and uh to the point where I’m actually somewhat a little relieved when I hit my limit and I can take a break for a minute.
J. Langley: You know, I think there’s been some papers and some other research actually about um how people deal with that. I mean, it’s such an overload of drive, drive, drive, drive because it doesn’t get tired. It gets lazy sometimes. Why is it that the engineers at work are not as fast like they cannot use the tools? Yeah, they can’t. Some of them can’t use these tools. Uh some of them don’t want to use the tools. Um some of them and again the some of the parts of it are depending on which models are available at which customer site. Some of these models I don’t want to touch either. um you wind up fighting with it more than you, you know, especially if you’ve used good things and and now you’re on something that you’re spoiled. Yeah. That might write a unit test if you talk to it in the right way. Sure. Um so there’s a lot of that. Um some of it is education, some of it is uh just getting your hands on things.
J. Langley: That’s part of why we do what we do here is sometimes you wouldn’t believe what I’m telling you unless I show it. Um I’ll call that the magic wand approach. Sometimes you go in and you’re like, “Hey, we’re doing this.” They’re like, they just dismiss it because it is so far outside of what they’ve ever seen that it’s just like, “Okay, sure.” Um, but if I told you that everything that this has written since, uh, actually, we’ll get to that in a second. Um, so what I’ve actually built to start with is, um, I can’t afford just to go buy a bunch of robots. They’re they’re not cheap. Um, I thought about moving up to like drones, but those are also not cheap, especially now that you have to buy US space drones that are like double the price and all this kind of stuff. Um, so I started off with what is something along the lines of working in agriculture technology, a tech kind of stuff plus AI that I can actually do? you know, what is it that I can build up and possibly get some revenue stream somewhere going and then use that to fund whatever else I might wanna want to play with.
J. Langley: Uh, so what I came up with, um, some of it comes from some old NASA Space Apps challenge stuff we did with using satellite data. Um, so I’ve done that before or tried to guide teams of high schoolers to do that, which is fun, uh, in a different way. Um, so trying to figure out can I actually uh can I actually estimate what a crop is going to do based on satellite imagery uh climate data other kinds of information that I’ve got all free um in a way that can actually generate alerts for a farmer for something that they may not be able to see right now. uh currently they do things that uh they may notice certain crop things that happen but typically it’s it’s almost too late for them to react to it. Uh so the whole point is what can I do early on? Um and starting off I’ve got a uh I’ve got two different optical satellites. Uh there’s two sentinel satellites that fly overhead um about once every five days. There’s a LANCSAT satellite that flies over about every seven days.
J. Langley: LANCAD is a 30 meter resolution. Uh Sentinel 2 is a 10 meter resolution. Combine that with a uh Sentinel one which actually is actually radar which is a different part of my life. So it gets interesting. Uh the uh optical satellites uh are affected by clouds. If it’s cloudy, it can’t see the crop. It’s cloudy a lot during growing season. in case you’ve never been here in like June, July, August, you know, it’s hot, it’s muggy, it’s it rains a lot. Um, radar doesn’t care about any of that. So, there’s uh one of the things I had it implement is actually a model. Um, I think it’s a I don’t know if it’s a boosted decision tree. Uh, I need to go actually look and see what it built. Um, it’s a way to estimate the visual artifacts that from the actual radar data. Um, so that was something I I handed it a paper that was written and said, “Hey, can you build this for me?” And it said,
J. Langley: “Okay.” Um, and it did. Um, and then I said, “I don’t trust it. Can you build a way to verify it or validate it?” And it did. Um, and it’s typically about 92% correct uh across, and this is a pixel by pixel correctness, not just a average. So, that worked pretty well. uh climate data um using uh daily precipitation uh temperature min max um there’s something that farmers use called growing degree days when you plant a crop you can actually knowing how hot it is and and all kinds of things like that you know when your crop should emerge uh you know when different stages should occur and I mean I’ve most of that was me talking to Gemini and learning a boatload about agriculture that I had no idea. Uh so it was it’s been a lot of fun from that perspective. Um but uh let me go ahead and actually log in to this thing. Um I had it use actually let me sign out and do it again uh and see if it’ll bump me over here.
J. Langley: Okay. Um off zero I had never used this before. I need a way to, you know, make sure people are who they say they are and I need an email address. And so it said, “Do you want to use Google?” I said, “Sure.” Um, and so it built that out. And then, oh shoot, with things make sure you’re It has to make sure I am who I say I am. Um, yes it’s me and yes it’s 99. The set one data is how often? It’s a roughly about every 10 days or so. It’s not as often. There’s only one satellite that does that. Um, so right now I’m using that kind of as a fillin between the others. Um, and on the few days a year where the Sentinel one actually overlaps with the Sentinel 2, I snag that and use that to to validate. So I should be able to estimate what Sentinel 2 is telling me from the data I have of that at the you know for the same day at the same place.
J. Langley: Uh so currently uh the other thing uh that it did was uh and I’ll walk through some of this but um this is a field um right next to where my grandfather used to have a cotton farm. So I know exactly where this is. I know my dad was actually driving by there yesterday. I had him on the phone. Hey, is it is that cotton or corn or soybeans? And right now it’s corn. Um, but this actually shows you temperature data daily. Uh, each one of these dots is actually a sent I’m not I don’t have the Sentinel one laid in here yet. So, this is all just the visual stuff over time. This is actually starting to ramp up as it’s this is this is actually something called NDI or farmers call it figure. It’s basically looking at the signal coming from the green wavelength of that. it knows how green that field is. Uh there’s a different thing that we can look at as well. Uh in DRE, it’s actually looking at the red edge.
J. Langley: So you got green, but then when parts of red start to blend in, that’s when you start to have uh either disease pop in or some kind of a pest management issue or something. The interesting thing I’m already able to do with this, I don’t think I’ve got a screenshot of it. Uh, a lot of times farmers when they go and they look at this, they’ll before season happens, they’ll go do soil samples about every 10 acres in this field. Um, and then they’ll send them off, wait about three weeks, get a get a return back. Then they know kind of what the makeup is, what nutrients they need to add, and all that kind of stuff. Uh, with this visual approach, um, I can actually show where the good parts of the field were and where the problematic parts of the field were and show that you’re not going to find most of that if you just do a grid every 10 acres. I can say, hey, on this field, you need to check right here for the bad spot and right here for the good spot.
J. Langley: Because it’s some of these actually follow kind of like underground water line, not water lines, but you can actually tell where some of the drainage is. It’s interesting. Um, anyway, that’s just talking about the whole uh the the kind of thing it built. Uh, I had it build a design system for me, pick colors. Uh, I wanted it not John Deere green because I don’t want to get sued. Um, I needed it to not look super techy. Uh, but I needed it to what I I think what I gave it was find the types of information farmers currently read and make it look like it would fit in with that. So, it’s looks familiar. It’s got earth tone colors. It’s not a white background. It’s close, but it’s more of a more of a earth tone kind of a thing. That type of thing. Um, yeah. Number for details. Nice. I got details on certain things. Uh you can flip between seasons.
J. Langley: Uh they let it go dormant last year or at least for this number of days. Uh you do a a custom selection of what days you want to look at. Um what I’m what I started I I thought I’d finish it by today. Um but I didn’t quite make it. I started it yesterday. Um which is uh there’s two items. One of them is something called a CDL. uh per county. Um the government produces a big TIFF file for that county and each pixel in that file is a 30 meter by 30 meter. It will tell you uh what crop was PR was planted or what the ground cover was in the previous year. So what I’m looking at is a farmer logs in and picks a field. I want to actually go pull the previous 10 years worth of data. Satellite all that’s free. That’s once a year. Huh? That’s once a year. No tiff file. Well, that tiff file is once a year and because you have like three crops in the same field and so you may get just one of them.
J. Langley: That’s a different color. They’ve taken into account that some people have winter wheat. So it’ll it’ll have a double crop that’ll tell you which crops it were in what order. Um and so it’s pretty interesting. Um and then the other part I’ve pulled uh that’s very close, let me see if I’ve got that up here. Uh, let me find my data file real quick. Uh, anyway, this is all the code. Yay. Um, you’re just serving this on your laptop. Uh, right now all of this is also everything I’ve got is hosted on Amazon. Um, if you look, this is uh this is actually crap, what’s it called? Cloudfront. CloudFront. Uh, so it’s pushing to an S3 bucket out front. Um, the API is actually hosted in a Lambda. So I there is no actual server running behind any of this. So when I loaded it, uh it you saw it was not instantaneous. It’s not but I’m also not paying for and and and doing anything to uh lock down a server, all that kind of stuff.
J. Langley: Um Claude actually was nice enough to build the entire Terraform thing that I needed to actually manage the whole infrastructure. Did you look at the code or have you looked at the code? Well, I’ve looked at I’ve looked at the code. How much have you modify the code? No. No. Anytime I need something modified. I actually this is we’ll we’ll segue in just a second. Uh I also found that one thing uh farmers use a lot is insurance actuarial tables. If a crop goes back, I mean that everything is covered by insurance and it better be because you’re going to go under as a farmer if you don’t have that. So that that covers uh the earliest planting date that’s allowed the last plant time you can har I mean there’s a lot there. Um so I just had uh actually that’s what I was looking for. I just asked it to go hey can you figure out how to pull that in so I can use it. Uh so my primary data set this is RMA actuary for 2025. Um I can see that and this is all broken down.
J. Langley: FIPS is actually a county code. So per county uh per commodity uh commodity name, insurance plan code, revenue protection, blah blah blah. Um I can actually see the earliest allowed plant date under that policy, all of that kind of stuff. They publicize all that. Yes. Insurance. Yes, that’s USDA type information. Um and publicly you can go download federal insurance. Uh yes. Well, the feds decide the dates and all that kind of stuff. And most of your private companies, they don’t have to pay to decide the dates. So, they can blame somebody else. Yeah. And so, yeah. So, um their plant dates themselves, but through the every five day satellite data you’re getting, could you also verify? Yes. Perfect. So the two things I need, one of them is um you can see the uh signal from the optical side start to go up for the green. You can als there’s also a signal that you can determine from the radar if it happened close enough because again some of these like cotton might be in the ground for weeks.
J. Langley: before it emerges and that gets I mean you want to talk about uh anxiety you know corn 10 to 12 days you know things like that are a lot quicker um the main reason I did this was if somebody goes and enters a new field I didn’t want to have to go ask them for 10 years worth of information for what what was planted when did you plant it I want to go ahead and go you know because a lot of it doesn’t really matter they just want to compare one year to the next so any of that I can automatically feel I do. Um, similar to what you ran into, uh, I took someone’s advice and started using, uh, clawed design. Yes. To build a slide deck, which it did. Um, and similar to you, I cannot for the life of me to get it to center this on every freaking slide. I’m like, can you just put it in the middle? No, it will not do it. Um, so I’m almost at the point of shipping it to Gemini and trying to fix it and getting it back.
J. Langley: Yeah, I’d actually slides HTML. Okay, we’ve had that. That was uh Tom, I think last No, he was I made mine at my last presentation last week with HTML. Okay, Tom was SVGs instead. Okay, that was Follow your design guidelines, too. We have a lot of trouble. uh most of my clients are on the design side and I am a hawk whenever it comes to design guidelines and you know a large may not care about the design guidelines but that’s something that I pay attention to with my work and so um it is much better to use code for that because it will actually follow your design tokens um and implement the custom I mean because technology can be all of your colors and um it also keeps that design system. So, it’s very easy the first time. Going to take you a little while to get to where you want to go and then every time after that, it’s much easier to generate anything and it’ll be the exact same every single time. So, you can tailor that presentation to any audience that you would need to.
J. Langley: Cool. Are you putting that in a still with this one with design? I jump on yours for a second. Yeah, go ahead. I’m looking for where it lets you do the actual U. Wow, this is a lot of prompts for me trying to explain to it what I wanted. The font design also takes up a lot of tokens to you. Oh yeah. Yeah. Really? It is extra. It is separate usage. how like you run out of regular usage kind of get a little extra out of time something I found that’s something I didn’t say during so when the initial game framework is that’s a lot of the basis for almost 3D game so I looked at uh one of the old magic super Nintendo games I’m like no this would completely plug into my framework and I was out of tokens and I just used claw design I’m like here’s everything I currently have here’s uh my magic 2 from the Super Nintendo put together a slide deck that I can feed in the clock to borrow and well I ran out of tokens at seven slides and got mad at it and stopped.
J. Langley: I’m like I was just so done with this. But this is where and similar what you’re talking about. You can have a load of design systems that you’ve implemented that each one is its own thing and then switch around if you want and things. But yeah, for sure it at this point with cloud design and I think this is true of a lot of the VI design tools. I would not say that they’re very good at following design system. Um they are moderately good at getting you about 90% of the way there with something possible. But as far as like technical design, I think cloud code gets a lot closer with better principles around what it’s actually using. Um and actually and it is much easier to actual right the uh well just a quick kind of what it did eyes on every acre every day because you can only see the field from the road if you’re driving by. You know things like that. You can’t scout what you can’t see. Um the actual information it pulled into this is actually correct.
J. Langley: Um and it’s actually um their fields um not just countywide or statewide stuff that USDA is always like wide area stuff. uh the different levels of uh actually it’s split it up into things you can affect today if you know about it or things that you actually have a minute to work on next week or hey this seems to be a a repeating thing um along with the context but you know interesting stuff um so so this could be your pitch guide for each yes that’s kind of what I asked it to do um the other thing I wanted to hop into real quick was what it was like actually trying to manage all of this work. Initially I’m called it Fieldfinder because I started off in a wholly different thing. I was trying to talk uh build a SAM 2 model that will actually you give it a center point, it does the outline. Um I’ve still got that in the plan somewhere. Uh but if everything I did initially um initial work I started off in kind of how I’ve seen others where I’ve got uh my documentation in one folder which I’ve got let me go actually pull that as well.
J. Langley: So, I kind of did what I’m used to to dealing with where I’ve got an architecture document, I’ve got an algorithms document, I’ve got some other types of things in different categories. Others have their own way of doing it, but I did this because it made sense to me. Um, and initially I had, you know, everything in there and everything in in in between is linked. I think we’ve talked about this maybe some uh in that a lot of times we do stuff agent-based or whatnot and it’s information that’s good for the AI to go build it, but as a human reviewer, I don’t want to be reading a boatload of text stuff. Build me a picture or whatnot. Um so I settled on using mermaid for diagrams which agents can also use you know read them like it’s a like it’s a diagram. Um system views uh package organizations what the data flow is. Uh that’s where you determine something wasn’t connected. Uh no I didn’t catch that until about a week ago. Uh so have uh you have you ever run into something with uh this is something that gets interesting.
J. Langley: I’ve worked with one one or two developers before that are so so fast at getting things done that it’s great when they’re going in the right direction. If they’re going in the wrong direction, you’re going to get the wrong place in a in a hurry. Um, imagine clawed code going in the wrong direction. So sometimes you ask any of these AI models something and then you you’re working on it a little bit. Wait a minute, something doesn’t seem right. You know, is can you go check this for me? And of course it goes, you know what? You’re right. I’m so sorry. Chewed up all your I’ve chewed up all your tokens. All this kind of stuff. So initially I had this whole thing built using kedro which is something I’ve used in the past for managing data and like that’s why if you look at uh in my data directory you will see a primary raw intermediate um you know feature model output model input um that’s just how I’m used to working with it.
J. Langley: So, I initially had it like that and eventually it it decided that KRO wasn’t the right way to go and it said what we’re going to do is we’re going to do uh we’re going to do lambdas on AWS and we’re going to do Daxter. Yeah, I don’t know if you’ve used Daxter before. Um I’m just going to skip it. We’re get close on time. Um Daxter is a way to manage pipelines visually and you can uh materialize your assets and see which things trigger other things and it’s really cool. Um, and it built that and then it built the step functions. And I’m at like last week I’m sitting there going, why do I have two things that both run pipelines? Mhm. Um, does my Daxter thing ever trigger my step functions what I have in AWS? Oh, no. Do I need that? Well, actually, you don’t. I’m sorry. I’m sitting here paying $5 a month to host a freaking Daxter thing and trying to figure the reason I came across it was Daxter is open source.
J. Langley: It’s very, very useful, very intuitive, really cool. um it does not have an authentication package. If you want it multi-user or anything like that, that is on you to figure out how to package it. So, I’m like, okay, now I need to put engine X in front of it or some other kind of thing. I’m like, wait a minute. I why do I have this? So, yes, um one of the issues I came across um in my issues list here, uh remove Daxter from production. stop it. Um, luckily it doesn’t get angry at you. Yeah. Um, and it said, “Okay, five minutes later, Daxter is no longer part of my production.” Um, I told it to keep the parts of it so I’ll actually remember later because my memor is not great. Um, so instead when I’ve uh I’ve got this thing uh in AWS and the fun part of all of this before I had worked with this, I didn’t know this was even a thing. I have never built a step function in my life.
J. Langley: I didn’t know what state machine I knew what lambdas were. Uh but I didn’t know this is a whole collection of lambdas where the input of one drafts the output of the other and all this and I can kick it off with one command. Um or in my case, put a rule in that kicks this thing off every day for a specific field. Um and it cost maybe eight cents. So anyway, interesting crap there. Um but what I had run into was I was trying to do uh the same kind of thing in where I’ve got my specifications that I’m going to ask the model to go build or the agents to go build, all that kind of stuff. But I had these in my codebase and every time I needed to change one, well, now I’ve got a whole another commit or a merge request or something uh and I was actually um it felt like I was slowing down the actual thing that needed to build. So eventually I wound up taking the entirety of how I tell this thing to operate and moved it to GitHub issues uh along with setting up a GitHub project where I’ve got linkages of things and you can actually kind of see the workflow as it happens.
J. Langley: Um and I’ve got clawed um I didn’t do it as a skill. I just told it to use the GitHub API um on command line. And now and in my agents file I have it where specs are GitHub issues. when you are ready to when you’re done, create a pull request and add me as a uh you know, put me in as a reviewer. And now I’m back into more of a flow. Uh bathroom straight down the hall if you’re looking and anybody else that’s new here, which are all but two of you. Um so that’s kind of uh how I’ve been running this whole thing. You can see this is just nearly all of these are I would say a at least a story level um thing that you would give to a team and it’s just mind-blowing how much stuff this is cor crossed. So then I I got curious and I was like, “Well, hey Claude, um if you don’t mind, uh are you familiar with a Cookamo 2 software estimation model? I have to use that at work when we go build bid new work and things like that um on other projects.” It said,
J. Langley: “Sure, I know what that is. um will you run this through that model um and see what the person estimate would be to build what I just built? Let me flip this over to a preview so you can actually see it. Uh currently I’m less than 30k of slack uh logical slack um third party with the SAM 2 um effective slack module breakdown uh between all the different pieces. is uh scale factors. This is kind of interesting. It’s smart enough to know when you’re building out one of these models by hand, it really cares a lot about the hard things. Uh like the thing I had I had it actually build an S build a model um using crap. Uh B the one I was talking about where I fed it a paper and it built the Oh yeah, there’s a lot of math in that. There’s a lot of satellite interface. Yes, it’s a lot of things. So, it’s not just a it’s not just a crud. Here, hold my thing and display it properly.
J. Langley: Um, so by the time it got down to the end of it, schedule, average staff size, it figured I needed eight people. Um, we’re at nominal effort for 133 person months. Um, and I’ve been working this on the weekend at $20 a month and just basically care and feeding of the no, don’t do that, go this way, no, try this, no, try this. Um, 17 months. Yeah. Is so if I go ask for money, that’s probably what I Anyway, if you want to know, it gives you the whole uh the SNAFEX estimator is the part where it uh random forest pipeline. um and how to do that uh all the different external data sources, AWS orchestration, you know, all of that type of stuff. But, uh the interesting thing for me is kind of if if the work I do daytoday starts pulling in more of this agentic stuff and more of the code generation stuff and actually builds this in, we’re going to be screaming so fast. Uh there’s if if we don’t also pull this in on the test side.
J. Langley: There you go. Good freaking luck. Done nothing for velocity. Yes. We are stuck in QA. Yeah. And and in fact we realized because of this we were always stuck in QA the entire time. Right. And we never automated anything. And the entire time we had developers they were still waiting on QA. So we had developers drawing code and QA was still testing it. That was our bottleneck. Right. Is it before AI? So you started using GitHub. He said he was ups. Yeah, I was I was actually reviewing um and then occasionally it would check something in uh that didn’t pass the pi test because of course it’s writing tests. It loves writing tests. Um uh sometimes it wouldn’t pass the rough check, you know, some things like that. And now go. No, really. Go do this. Um, lies to you. Yeah, it does. Um, I passed.
J. Langley: So, the two things the two things I need to do with this one. Next one is to figure out how do I actually instrument an automated test for what I what it’s built you know in a way that makes sense because right now I am hands-on spinning up a lambda testing it first before I it goes production all that kind of stuff um I need a staging and a production area and I you know all that kind of stuff which I’m sure it can do um and then the next thing I need to work on is the upfront stuff um actually writing the specs and actually figuring out what the next thing is and what’s the next thing we want to build and um luckily for me that’s where I actually think the fun part of work is. So I am actually know I’m going to do that you know um yeah have you run into yet? It’s obviously not this project, but others for the QA testing where I do have it. I’ve had some agentic setups get so hung up on testing that I’ve had trouble getting it back into production that I’d actually have to to start adding a because I am doing the merge requests.
J. Langley: I I say there’s a net plus uh plus amount of code you can do in this merge. If I don’t do that, it’ll write tests. It’ll be it’ll be 3,000 lines of tests and 500 lines of feature. And I’m basically like, okay, that’s great. So, whatever your merge is, you need to do half of that. Suddenly, those tests start disappearing out of nowhere and have 45 more tests with no actual. Yeah. And they’re just testing if the function returns what the function said it’s going to return. Is the string the string that we made a constant? Yeah, it is. It does. I’ve actually found that like just give it guidelines say that’s 80% coverage. Yeah. And then that’s that’s what has worked at least qualitywise. That’s like worked best for me. I don’t really care what the tests are. It you know it’s still got to check if you’re saying the coverage there’s tools that like check the coverage and they’re for AI.
J. Langley: So u it has to it still has to make those pass. Um, and um, so for all I mean that’s that’s the one thing I think is like if you’re going to do something to to increase the quality like put that in as a guardrail, right? Say 80% test coverage and then that will at least I mean even if you have like 2,000 tests are on time close agent come in at the end before I look at it. It’ll forget what those guidelines are. But if you have an adversarial agent set up and you’re like, “Cool, I’m looking for best-in-class like implementation practices. No dead code. Tests need to test actually valid stuff or they’re useless and I don’t want them. Let that guy run at the end of every session when it says it’s done.” It’ll come back and be like, “There’s dead code. Those tests need to go away.” You get back to Google agent like, “Yep, it’s right.” Yeah. Yeah. Let’s go ahead and let’s go ahead and do that.
J. Langley: I’ll get off of this. That works. 10 to$15,000 mark and a half years. Well, the the two things that definitely the vibe thing wouldn’t have caught. One was it was building up two totally similar stacks without connecting them and whatever. The second thing, and this is something I’ve run into a couple of times, one of the pipeline things it did for pulling data from the climate piece, kicked it off, and it’s pulling like 10 days worth of uh climate data, and it’s running for like an hour. I’m like, this is obviously slow. U plot has no guess of time, you know what I mean? It’s just like whatever. And come to find out it was doing like one socket at a time and then waiting on that time out before it would do another thing, you know. And after I identified it said, “Hey, this takes too long. It should be within this range.” Is like, “Okay, I’ll optimize it.” But it had no idea it needed to without somebody going and telling it. But back to David for this part. It’s hard to know what time it is. Well, it’s not checking so it doesn’t know, right? \