Slightly Caffeinated

TJ (00:00)
Yo, welcome back to the slightly caffeinated podcast. I'm TJ Miller. So Chris, what's new in your world,

Chris Gmyr (00:04)
and I'm Chris Gmyr

Yeah, busy week. It's a short week for me. I'm taking tomorrow off because we're going camping with the scouts. So just feeling like I'm running around and doing a bunch of things to prepare for that because I'm the camping chair or whoever is in charge of camping setting up. So like I get the campsites all bought and paid for, a bunch of the planning stuff done. I build a trailer. So I have to coordinate with a bunch of people and get some stuff kind of.

TJ (00:16)
Yeah.

Chris Gmyr (00:34)
last minute to make sure everything goes smoothly. So about halfway through. So I usually take the day off of the day that we go just to finish everything up and make sure all our stuff is packed. But yeah, going pretty well. And so far, knock on wood, the weather looks pretty good. It looks sunny in between like 40 and 65 degrees for the lows and highs. So it'll be a little chilly, but I'd rather have that than camping in like 80 degree.

TJ (00:47)
Yeah.

Chris Gmyr (01:04)
blazing hot sun so no problem.

TJ (01:07)
You can always,

like, I always look at it this way. You can always put on more clothes. You can always get in, like, another sleeping bag, you know, but, like, there's only so much you can do when it's hot.

Chris Gmyr (01:12)
Yes. Yep.

Yeah, yeah, exactly. So just plenty of layers. I'm getting a bunch of firewood, so I'll just get a couple bundles more, because that's been nice in the past, where we just light a quick little fire in the morning while everyone's waking up, waiting for breakfast, getting their coffee. And the kids just thaw out a little bit before we get going for the day on Saturday. But yeah, it'll be a lot of fun.

TJ (01:34)
Mm-hmm.

Chris Gmyr (01:39)
Yeah, it'll be a good time. So just kind of prepping for that, doing a bunch of like project planning at work, tinkering with a little AI stuff. Hasn't been like as busy for AI stuff this week, just because I got other projects to work on. But yeah, hopefully get back to that a little bit more next week and in the future. yeah, all is good over here. How about you? What's new in your world?

TJ (02:02)
Sick, man. No, last week has been just, it feels like it's been a little bit of a whirlwind. I've been wrestling with contractors trying to get stuff done around the house and that's been frustrating. They did good work, but man, was it three attempts before they actually got here to do the work? So three different scheduled appointments and not able to show up.

Chris Gmyr (02:18)
Yeah.

TJ (02:29)
annoying because like I need them to do work so I can do work and then I can get other people in here to do work. So it's like, it's just like holding up this big train of stuff that I need to get done. So luckily that's all taken care of now. Works been like pretty cool. got to work on some AI stuff, had like a major kind of like breakthrough with the architecture of like the click house stuff.

Chris Gmyr (02:37)
Yep.

TJ (02:53)
⁓ we're now in the stage where we're implementing the reads and like we went into it with one strategy and that ended up spilling into like a lot of work and just like felt risky. so we're taking a little bit more of a surgical approach and just like updating things. Like we were gonna do one big sweeping update, I guess to be a little bit technical cause it's kind of interesting. What we're going to do is we have a like request repository now that's like

is the access layer for doing anything with a request logging. So we have one that's MariaDB. And then I was going to create a interface for it, then also create a click house one, and then just use the IOC container and swap the implementations out. That would be cool, but there was like, I ran into untested code, and it just started spilling out into this huge, huge project.

like, well, maybe if we take a step back and do a more surgical approach, we can actually like just start getting things done. And like actually like shipping changes. So that's what we did. Took a step back to that. like been refactoring all of the like, the queries into actions, and then handling any sort of like MariaDB, like co-entering or

like dual, like dual entry or like swapping between the connections inside of these actions. there's like, basically each action has like two methods. There's like from a, like from Maria method and like from a click from click house method. And then I like in the invoke method, we just do a little bit of juggling to like determine like what methods need to be called, but super simple approach and like actually making forward progress on it now, which feels good. Cause it's kind of like,

stalled out, but yeah, I got to work on my first AI project too, which was like a ton of fun being able to build with Prism. One of the cool things is someone had contributed Boost AI guidelines to Prism and they were pretty extensive and like took up, it was like 4,300 tokens or something. So it was like a little bit too much.

Chris Gmyr (04:52)
Mm-hmm.

TJ (05:01)
But Pushpack came in for the win, shrunk them down and dude, I was blown away. Like I spun up a Laravel app and then I installed like Boost and it like hooked into the Prism AI guidelines. And I was like, yeah, I need to like build this thing with Prism. And it like built the command, built like, I was like generating a report.

of like a whole bunch of things. like built out the prompts using blade views. It like built out the schema for structured output, use structured output to like generate this report. Like did all of the things. And of like very, I don't know, I only made a few tweaks, but I was like very impressed with the quality of everything and how well it knew how to use Prism using those like AI guidelines.

So I don't know, man, that was like such a cool experience to like watch AI like correctly build with Prism was just like so cool.

Chris Gmyr (06:08)
Yeah, that is really awesome.

Yeah, just when everything gets dialed in a little bit more, and the further you get away from the vanilla installs of all these things, it's like you can just see all the pieces falling together and everything just getting a little bit more solid for each iteration that you do. So it's really cool to see that evolution of these AI changes and just all the updates that everyone is putting out too has been awesome to see and take a look.

TJ (06:36)
Yeah. Yeah. So it's been cool to like, I'm actually using AI at work. I'm using Prism. so we're like, we're slowly making that transition. had that, like that job title shifts into like, principal engineer for AI. like, we're starting to get into that shift of like getting out of some of the other like app world projects that I've been working on and getting into like actually AI stuff now, is, is pretty cool. So, definitely appreciate that. Yeah.

Chris Gmyr (06:58)
And you can take it with Prism too. So that's even cooler.

TJ (07:01)
Yeah.

And I can't wait to see like, the more I get to use Prism, I can't wait to see what that means for Prism. You know, like, I don't know what it's going to expose to me at like, maybe I'm doing some like weird stuff with the API and I just like haven't noticed it because all my time is spent in Prism rather than working with Prism. like, you know, that perspective shift of like actually spending more time building with it.

It'll be interesting to see what comes out of that.

Chris Gmyr (07:33)
Yeah, definitely. I can see a lot of things being tightened up. Not that anything is necessarily bad right now, but just that finesse that goes into it. And Taylor always talks about that with the things that he builds and the APIs that he creates. It's standing off the rough edges.

TJ (07:53)
Yeah, yeah, like we can get into some more when we talk about Prism later. But there's there's a few things I'm going to be making some shifts on for a little bit. And I think I'm going to maybe push out the 1.0 release a little bit until I spend some time like. I want to spend a little bit more time building with Prism for a little bit to like see if anything kind of like shakes out beforehand, but.

Yeah. Yeah, man. Yeah. You want to talk some coffee?

Chris Gmyr (08:13)
Yeah, that makes sense. That's good.

Yeah, let's talk some coffee. So.

TJ (08:19)
because I am drinking

an amazing cup of hologram right now. I'm like loving this, like fresh, fresh, like I made a fresh press right before we sat down. So I'm just like living for it.

Chris Gmyr (08:31)
Nice. Yeah, I'm almost done with my cup, unfortunately. So I don't know, might make another one. We'll see how it goes. But yeah, the other day, sat down with one of my coworkers, one of our past coworkers. We do like coffee and coworking about once a month or so. And...

didn't really know what I wanted from the menu, but he got a red eye, which is just drip coffee with an extra shot of espresso. Like, that sounds good. But when I went up to the counter, I asked for the large, so it's like a 20 ounce or something like that. I'm like, is it still just one shot? Because some people still just put one shot in, or for larger volumes, they'll put in more. So I just wanted to make sure.

TJ (08:56)
Love it.

Chris Gmyr (09:12)
And the barista was like, no, it's still like one shot. I'm like, can I put two shots in it? She's like, yeah, you can totally do that. And that's called the black eye. OK, cool. I'll take one of those then. And then we did some work. We chatted and hung out for a couple of hours. And then when we left, we ended up getting two more, like one each, ⁓ two more black eyes. So I took that home. And it was already like 1130, like almost 12. So it's pretty late to be.

TJ (09:21)
Mm-hmm. Yep.

Yeah.

Chris Gmyr (09:40)
drinking coffee like that, at least for me recently. But I kept going, got busy in the afternoon and was like, damn, like I have not finished this yet. It's like four o'clock. I got like half of it left because I just get busy with other things and don't really think about it. So like polished it off at like four o'clock or so. And man, I was wired the entire night. like, I got up, like did a couple of things like.

TJ (09:49)
Ha ha ha ha ha.

Chris Gmyr (10:06)
I don't know, just tried like a meditation and stuff. Like nothing was working. So it was about like 1230. downstairs, like got some water and went back up stairs and just tried my best to fall asleep. But yeah, I think I finally fell asleep at like one o'clock in the morning. So ended up missing my workout session on Wednesday morning. And yeah, I'm like, okay, well, can't do that again.

TJ (10:31)
Yeah.

Chris Gmyr (10:32)
Gotta finish it off earlier, get like a much smaller like to go cup. So it's like out of my system. which is weird. ⁓ cause I feel like years before, I don't know, maybe like five, eight years, I'd be able to get like espresso shots, at night at the old place that we used to live at, there was like a, a wine and coffee bar. So like during the morning it was only coffee and like, ⁓ coworking space and at night.

They had a little wine bar that was open, but at night the coffee bar was still open. So we'd go there, get a little bite to eat, listen to some live music or something like that, and I'd get an espresso or two and come home and pass out and be totally fine. But yeah, something shifted in the last handful of years and apparently cannot do caffeine super late anymore.

TJ (11:12)
Mm-hmm.

Oh man, that's uh, yeah, it's funny, man. The joys of ADHD, like caffeine affects me a bit different and like I can slam a coffee and go take a nap. Like, uh, it can just happen. Um, so I, I normally, I normally re-caffeinate at like two, three o'clock in the afternoon just to kind of get through the tail end of the day. But, uh,

I really can't go too much later than that because like I will.

I'll still be able to fall asleep, but it affects my like quality of sleep. Like if I have like too much caffeine, I find myself like waking up frequently throughout the night. So I get like really broken sleep, but I have no problem like falling asleep.

Chris Gmyr (11:57)
Mm-hmm.

TJ (12:05)
But I do have to watch it now that I'm getting a little older.

Chris Gmyr (12:08)
Yeah, totally the joys of aging and all the shifts and changes.

TJ (12:14)
Yep. Yeah. Cool, man. I feel like we're about to talk a whole bunch of AI, and I'm really excited about it.

Chris Gmyr (12:20)
Yes, I got some AI workflow tweaks to talk about and some things that I've been tinkering more recently with and just some other, I don't know, things that I've come across, which has been pretty cool. So yeah, just wanted to share some of that for a few minutes. First, I message you on the side and on Instagram, I've been seeing this AI Black Magic ad. It used to be like AI.

like chat GPT something or other, but made it a little bit more general. And I've been seeing it for a while. It's all these different general writing rules, different ⁓ prompts that you can put in, different agents that you can download and reuse. And I kept on seeing it, kept on seeing it, and definitely got influenced. Then finally, the other night, I pulled the trigger and went for it. It was like,

TJ (13:06)
I love that.

got influenced.

Chris Gmyr (13:10)
influence

and pulled the trigger. It was pretty inexpensive. I think the base plan and I got some other template package as an add-on and it was like 40 bucks or something like that. But the base thing is like 27, I think. I'll put the link in the show notes. immediate boost in quality and productivity from it. I was pretty amazed. I've only used a couple

prompts and some settings so far. So I think the biggest thing that drew me in is they have this prompt that you can either put in like one off chats or I ended up putting it in the general settings in Claude. It's like all these general like writing rules to make it more like human like and easier to read and a lot more natural. So they give you this whole list of

common AI words and phrases and punctuation. And it gives replacements for those. So for example, don't use an M dash, because that's a clear sign. It's AI, which is, I don't know.

TJ (14:16)
Mm-hmm.

Here's

the thing, man, I actually use dashes all the time in my writing and like hyphens and stuff to like...

do things. like, I hate that it's this like AI thing because I love using dashes.

Chris Gmyr (14:32)
Yeah, yeah, and I maybe not like dashes as much, but definitely like hyphenated words. Because I know another tool that I use and have used for a long time is Grammarly, which is fantastic for writing. I feel like I've picked up enough by using Grammarly and like common replacements and changes that I've had to make to my writing. It's like, I just do that now for dash words. I just.

TJ (14:44)
Mm-hmm.

Chris Gmyr (14:57)
do that. So it's like, well, it's not really an M dash, just things that I do now. So I don't know. M dashes are little bit different. But what's really cool, and I can go through a few of these, let's say they give you expressions and phrases, a couple of punctuation marks and things like that. So one is like labyrinth.

TJ (15:04)
Yeah.

Chris Gmyr (15:23)
I don't know if an AI has ever said labyrinth in a doc that you wrote. But the replacement is a maze or a puzzle. So when you're writing as a normal human, you'd be like, oh, it seems like going through this project was more of a puzzle than we realized, where it would say labyrinth. So it will go through, and there's, I don't know, like 100, 200 swaps in here. And it'll say.

TJ (15:27)
⁓ for sure.

Chris Gmyr (15:51)
After you've write something, that's going to be like the regular AI output from that. Every time you write something before you present it to me, review your writing, scan for this text or phrases, and then rewrite based on the corrected version or the replacements that best fit what we're trying to write about. So.

It's been fantastic. I mean, even paying like the 27 bucks for like this one prompt for your like general writing rules, I think is worth it, at least for me. And I was like tinkering around with ⁓ some like, I just wanted to try it out and see what it would do. So I know we'll get into this, but Claude released skills recently. So in a separate chat before I made these changes, I said, write me a blog post of

I don't know, 500 words on the new skills launch. And then I added, and it was decent, but it was definitely like, you know, very like one-shot AI generated. And updated the global preferences in all these like writing instructions in the cloud settings and said, went back to the chat, I said, I've updated my global settings now with a bunch of like writing instructions.

TJ (16:49)
Mm-hmm.

Chris Gmyr (17:07)
Do you see this? Yes, it said. And then I said, please rewrite this blog post with all of the changes that you see in the new writing style and instructions. And it did that. It showed me all the highlighted words and phrases that it swapped out and then corrected. And immediately, it just sounded a lot more human and a lot closer to what I would have written. Then I moved that into the project that I have for my blog, not that I've

written too much on there, but it has its own set of writing instructions. And I adjusted it even further from there. And it sounded a lot more like me. If I was going to publish that, which I don't think I am, I'd update it a lot more. But it was so much closer than what the one shot without these writing instructions were. So I think I'm going to tweak it a little bit more.

based on how I write, especially for casual versus maybe work, like technical writing. But I think once I get that a little bit more squared away, I'm going to take these settings and then put it into my global Claude MD file for writing within Claude code. Because there's a lot of research or first drafts of technical documents that'll have Claude code write with me.

in the actual project that we're talking about. So at least these things will then transfer into that writing style before I bring it over to like Cloud Chat to really like tighten everything up or combine resources or something like that. So yeah, it was just like a pretty cool experience that just, again, like you were talking about with the Prism guidelines, it just snaps into place and it's like, it just feels so much better.

TJ (18:48)
Yeah, there's like, this seems really similar to some things that I use already. So when I have it like, write content for me, I have a like writing style guide that I've generated and I've talked about it before that like kind of has, it's a style guide to like write similar to my writing style. And I did that by like feeding it all of my blog posts and some other writing that I've done.

Chris Gmyr (19:04)
Thank

TJ (19:13)
and then said like, fill out the style guide based on that. And so like, when I have it right for me, I feed you this style guide. So it's very similar to like some of these things. So I'd love to like take a look at these prompts and see what I can like integrate into like my existing workflows, because I'm already kind of doing that. So when I have it work on like Prism documentation, have, I published it for contributors also, but it works really good for AI.

like a documentation style guide for Prism. like, I love to see about like adding similar things into like that workflow too, just to kind of get like a little bit more natural writing out of it. But like I've, I've worked with StunSpot and Collaborative Dynamic prompts for a while for kind of on and off for years. Like one of the ones you'll see me use a lot in like demos is Nova, which is this like persona prompt from these guys.

And I've seen just the power of these, what some of these prompts can really shift and guide the output. yeah, for like, to see, you sent me the before and after on these blog posts, very effective prompting. I think it would be worth the money. So yeah, I'll probably end up picking this up and try integrating it into some of my things.

Chris Gmyr (20:24)
Yeah, yeah, that's awesome. And they have like tons of stuff in there too. Like I've only just scratched the surface and tinkering with it, but it seems like a pretty decent product to grab, especially if you're doing a lot of AI work and prompting and just wanting to tweak it a lot more in your.

TJ (20:41)
Yeah,

we can touch on it a little bit more, but I think like that kind of stuff's primed for like Anthropics new skills stuff, right? You have like, you could be like, yeah, here's like a writing skill. We'll talk, we can talk about more about that in like a minute too, but I've been playing with that a little bit more. And so like, I think like that kind of stuff is like prime for a new like skills platform.

Chris Gmyr (21:02)
yeah, definitely. So we will get to that. A couple more quick things that I'm working with too is Whisper Flow. It's whisperflow.ai. And this is like a voice to text like AI that polishes what you say too. So it'll get rid of like the ums and the ahs and the pauses. Or if you like double up on a word or something like that, it'll clean that out before it

TJ (21:05)
Yeah. Yeah.

Chris Gmyr (21:29)
actually either copies to your clipboard or injects it into whatever sort of application that you have. So I just signed up for the trial yesterday. And I think it gives you like 200 words that it'll transcribe for free during the trial. I've already gone through like half of that. Yep.

TJ (21:50)
Yeah, it's

like 2000 words per week.

Chris Gmyr (21:54)
Oh, perfect. Cool. So I'll probably run out of that pretty soon, since I already went through half just yesterday and got a while left to go. But it's been fantastic. So I've put this into emails already. They have not only a Mac program, but an iOS keyboard application. So it does that locally on your iPhone as well. So I've done a couple of text messages with it to test it out.

I put this into Cloud Chat. I put it into Cloud Code. And yeah, it's just so nice to speak something. And they have a couple of options for AI coding and vibe coding in there too. So it'll pick up, if you say, like, at in a file name, it'll try and pick that up, or like an agent, or something like that. So it's really cool that they have those options in there. And.

Also a system wide shortcut. So it just needs to be like running in the background. And then there's a couple of ways that you can invoke it. So I usually, I have it tied to control tilde and that will start and stop the recording. So I can basically like hit that talk for as long as I want to, and then hit it again. It'll stop it. There's like a slight little pause and then I'll be like, boom, right in the application for all the texts. And then I can add to it. I can clean it up. can move things around if I want.

But it's just been so much faster than typing some things, especially since just bouncing around to different apps. I don't know. I just found it pretty cool. And also, it feels more futuristic. You're not really doing the code either, and then you don't really even have to type. You just speak it, and it goes. And it's just super cool. That's it.

TJ (23:40)
You

I'm a hundred

percent trying this. yeah, this is cool, man. Like pricing seems pretty reasonable to you. but man, I love, I love doing this stuff, especially like I've been going for walks a lot lately, like just being able to like go out for walks and like voice memo and like pipe it through this to get, you know, I've done that before where I've had like workflows that like transcribe it and then like output it into like formatted notes, but

I like that this also does like grammar fixes and stuff on the fly.

Chris Gmyr (24:12)
Yep. Yep. Yeah, and they have, like I said, a lot more tools for developers in here. Context Aware, it'll do like Claude stuff. It'll do stuff in like Cursor and Replet and a whole bunch of tools that it's already specifically built for, which is just, I don't know, it's really cool. Especially for the lowest base package, like the Pro Flow.

$12 a month like billed annually, like that's not too bad. Especially if you're doing a lot of prompting and a lot of projects with AI or you just typically type a lot and don't want to type a lot. It's sometimes a lot easier just to just speak what you want to do. So yeah, I've been pretty happy with it so far. Probably end up buying it and then working it into my workflow a little bit more.

TJ (25:01)
Yeah, this is interesting, man. I'm

captivated.

Chris Gmyr (25:04)
And I've tried using a couple other open source options, but they didn't really work for me. I don't know why, maybe, I don't know. A lot of these things you have to kind of keep up on, especially for the local models. But I tried a couple, didn't work. I have a subscription to SetApp as well, which has a whole bunch of different applications that you can install basically for free in the subscription.

plan that. Yeah, it's so nice. There's so much cool things in there. I'll have to put a link in there to set up too. But there's a few speech to text apps in there. And I tried one of them out that had the better reviews, but it was just janky. It wouldn't always keep the settings for the microphone that I wanted or the model that I wanted. And I know. I just tinkered around with it for a day or two and then

TJ (25:34)
love set up.

Chris Gmyr (26:01)
It just wasn't working consistently. So I saw like whisper flow on something. And I'm like, OK, I'll try it out. I feel like I've heard of that before. And just like immediately had success with it. And it just worked out of the box and what I wanted to do, especially for the invisible hot key usage. It was just super nice. So probably just stick with that and then see how fast I can rip through the trial and just spell the trigger after that.

TJ (26:27)
Yeah, yeah, man, I'm definitely gonna play with this. This looks really cool.

Chris Gmyr (26:31)
Cool, yeah, and then lastly, just real quick, not really related to AI, but I've been playing around with iOS shortcuts a bunch to just try and automate either things like on the computer a little bit more that need a little bit more like either Mac OS compatibility or iOS. And I've always not found great success with logging water.

between water bottles and going to the gym and stuff like that. It's something that I wanted to do, which seems pretty easy. But doing it through my fitness pal is kind of janky. When I had Fitbit a while ago, it just didn't give you either added up everything in one bucket. I'm like, did I add the thing at lunch or not? And it didn't really, I don't know, work out very well. So looking around, and there's a bunch of paid apps too, which is kind of weird. I'm not going to pay for an app to just log my water.

But I was tickering around with shortcuts and you can actually give it access to your health data in the app and be able to add stuff to there. So one little shortcut that I made was to log my water bottle and you have it set up like a input prompt and then that feeds into I think it's ⁓ like add health entity or something like that. So you take that output from the input

which is like maybe 18 ounces or six ounces or whatever. And then I'll feed it into the health app automatically and just log it. So then you can go into the health app and you see all the individual logs throughout the day. So you can see 8 AM versus 10 AM versus noon and go through like that. So I'm like, that's all I wanted. And then it was free for like five minutes of work. And it was pretty cool. So.

TJ (28:15)
Yeah, I was shortcuts

like I got really into them for a little while and they're fun and really pretty powerful. I haven't played with them in a while, but yeah, you can do some really wild stuff.

Chris Gmyr (28:28)
Yeah, and people have some crazy big shortcuts and some things that I would never use. So some people have it grab your current location if you're driving to work. And then once you get there, it'll text your spouse that you arrive safely, automatically. Same thing when you're leaving work. After 4.30, when I leave this location, then text my spouse that I'm coming home or something like that.

TJ (28:44)
Mm-hmm.

Yep.

Chris Gmyr (28:53)
It's just cool that you can do that. Like, I don't know, I would just do that manually because I don't know what if you're walking in and out of your office like multiple times a day like that'd be pretty annoying. So I don't know. I just have it for a lot of like simple things that I do a lot. But one that I another one that I set up is just like reconnect my AirPods because sometimes I connected to like one of the laptops or like Apple TV or something like that. And I don't have it like automatically connect based on

the audio output because sometimes I want to be just long-term connected to a device without having to switch. Like if I get a phone call, I don't want to switch from my Mac or iPad back to the phone to then answer the call, especially if I'm just going to ignore the call if it's from like a... I wanted to get rid of that for...

TJ (29:38)
yeah, the device juggling is really annoying sometimes.

Chris Gmyr (29:43)
the Bluetooth devices. So I just set up a quick shortcut on the home screen to just reconnect to AirPods. And that's been probably the most used setup so far, besides the water now. So yeah, just really cool things that you can do with that, and relatively easily, assuming the app that you want to use supports shortcuts. So yeah, something else to ticker with and mess around with.

TJ (29:54)
Hahaha

yeah, for sure, Sick. continuing on the AI front, man, there has been a slew of new things that have come out recently. ⁓ yeah, dude, we touched on it a little bit, but like anthropic skills. I've been starting to play with them. haven't, I, I've set it up. I haven't had a lot of time to play with it too much yet, but I ripped out.

Chris Gmyr (30:19)
Yeah, unphysically.

TJ (30:33)
my agent, was getting some like, still just getting and experiencing like weird things. And I sent you a couple of screenshots too of it, like weirdly trying to invoke the Claude agent through command line prompts that don't exist. And it was like, catting instructions to temp files to like feed the agent context. I don't know what the heck was going on. It was just like, I would always get to the end result fairly successfully, but it would do some like.

really weird stuff along the way. So I ripped that out. I like re updated my Claude MD file. Cause I even though I'm sure Claude's like base, the Claude codes base prompt has information about like how to use agents and how to use skills. I still explicitly write something to my Claude MD file for like when to reach for this skill or this agent, you know,

Chris Gmyr (31:28)
Yep.

TJ (31:28)
So I put a

little blob in there about like when to reach for the skill, even though the skill already has that. And then now the skill has like all of the PHP development and like Laravel development, like standards and everything. And then I think I'm going to split one out into like a PHP testing one with like all of the like pest specific stuff. But I'm going to try using skills instead of agents and see what happens.

I know that there's like the whole context thing is what's advantageous about agents, but I'm not completely sold on that yet.

Chris Gmyr (32:03)
Yeah. And I set up some skills the other day. I basically had like Claude scan the code base, the agents that I set up, and then the custom commands and said like build or suggest skills that we can build based on what we already have set up in the ecosystem over here. So I think I'd spit out like 10 options. I built five of them.

which I had it try to build its own skills, but it was missing a couple things, which is kind of odd because it was pulling from the live documentation with examples and stuff. So it took a couple of times to go through that, which was kind of annoying because they weren't loading. And it was because they didn't have the front matter, like YAML stuff in there. It was just kind of like plain markdown files. So I'm like, no, we're missing something. It's still not loading after multiple refreshes.

TJ (32:36)
Hmm.

weird.

Chris Gmyr (32:54)
I'm like, check again, seems like we're missing something. And then I finally found it and replaced a bunch of things. Also, skills are a little weird because they have to be in directories with like a skills MD file. I don't know, at least that's what worked for me. But that.

TJ (33:09)
Yeah, they've

got to be in a subdirectory, but that's because skills, like you can also like bundle other things with skills, I think. Like I'm not, I don't know, cause I haven't had a need to do it yet, but like, guess you can like also bundle scripts and stuff along with skills and that skills are meant to be like bundle-able. Whereas like agents really weren't, but like skills you could like bundle into a package and distribute that way.

Chris Gmyr (33:37)
Yep. So yeah, it took a little bit of trial and error, but we got through it. But one that I did tinker with was an open API workflow. Because in this repo, we do API first. I have a bunch of architecture and API first documentation and some workflows within that. So it built a skill referencing all these other docs that we have in the repo. And then I said, just build a simple help

a Hello World endpoint, and it just went through and did it. It adjusted the OpenAPI spec. It ran spectral for linting, added tests, added authentication based on similar patterns that we have for other endpoints. And it was just like, wow, this is fantastic. Yeah, it's only a Hello World endpoint, but it did it really well and really fast compared to what I've seen for agents.

So there's definitely a lot of use cases for this. And what's also cool is that you can extract these skills that you have locally and put them into Cloud Chat online. If you enable skills within the settings, you can upload or add additional skills. So that's really cool if you have something locally, I don't know, like a local blog writer or something like that. You could extract that and then put it into your

web chat and have the same output in both locations. just the whole plugin and skills architecture is really extensible, which is pretty cool.

TJ (35:08)
Yeah, like just looking through some of the skills that they like provide by default in like the cloud desktop app is like they have an MCP builder, a skill creator, a slack jiff like creator, theme factory.

like all sorts of interesting things. So like that, kind of gives you some ideas for, you know, skills that you could create too. So, I don't know. kind of, it's just thinking about like how I was using the agents. It's like, no, this is like a skill. Like the skill is writing PHP and Laravel code, right? Like the skill is writing pest code. So like putting together skills for those things. And then I think like,

you could also put together a PRISM PHP skill, right? Where it's like, it has all of the like...

basically the same thing that you'd provide for boost guidelines become your prism skill, right? So, I don't know, I'm gonna definitely like invest into skills over the next couple weeks and see what happens.

Chris Gmyr (36:08)
Yeah. But yeah, lots of, lots of cool things that we can do with skills. So I'm hopefully going to be able to have some more time into next week. Get into that a little bit more, see, see how it works out and maybe drop a few of my agents as well. So we'll see.

TJ (36:22)
Sick. So we also listed down here, Claude Code on the web. I have not had a chance to play with this yet, but...

Chris Gmyr (36:31)
Yeah, I checked it out, but also haven't had a chance to play with it. Basically, you connect to your GitHub or wherever your code is, and it can just run on the web, on the fly with your instructions. And it can make PRs and all the things that you would expect Cloud Code Local to do. So yeah, going to tinker with that a little bit more. But I think it's nice that we're starting to get into a spot where you can have similar tools.

any platform or system that you're on. Especially like, I don't know, you could do like cloud code on the web for Prism bug fixes or like handling issues or something like that. You can just like tap away, you know, in the cloud app on iOS and just, I don't know, handle stuff like that or make a code change or have a PR waiting for you when you get back to your desk or something like that. Just like some really cool opportunities there.

TJ (37:21)
Yeah, no, I think that's like really very interesting. And kind of like following along on that, like guess Claude desktop is now generally available. I didn't know it wasn't generally available. I've been using the Claude desktop app for a while.

Chris Gmyr (37:36)
Yeah, yeah. I didn't realize that either, but I think now that it's just been more stable and they've been pushing out a bunch of updates with that too recently, it's just, I don't know, tagging in is generally available. So I've been using it for a while as well, and I think it's been fantastic.

TJ (37:52)
Yeah. Yeah. I love it. I use it all the time. there's not, I've got like a couple of MCPs registered with it too, but, I use cloud desktop, like all the time. So it'll be interesting. It'll be cool to like upload skills and be able to extend it. Cause there is stuff that I have been using with cloud code where I'd like to use the cloud desktop app, you know, or something on like iOS and be able to have it do stuff that,

I think there's going to be a few skills that I end up uploading to Claude to use.

I wonder if...

I wonder if there's like a skills API. Hmm. Makes me think of like.

If you can upload skills, I wonder if they have API access to the skills so that you could use skills with Prism, Hmm. Interesting.

Chris Gmyr (38:34)
Hmm. That'd be cool.

Nice. Yeah. And then let's see, the last thing that we have on here is a ChedTPT Atlas browser. This was just launched, I think, yesterday. So I haven't dove into it yet. Haven't tried it, but listened to a couple of things and read a few things about it. And it seems pretty cool just that it has immediate access to any contacts that you're using and can automatically go through.

the webpage that you're on. So I've heard examples of people analyzing Twitter posts or someone I was just listening to uses Riverside to record their podcast as well. So it taught it via a prompt of how it edits the podcast. So all the settings that you need to push and different ways to export it. So you have a fully edited and fleshed out show.

So it's like, that's pretty cool that you can just automate all these things via the browser. Pretty hard to do before. ⁓ the Chessy BT has some like lofty visions of like, ⁓ here's this item that I want to buy, like go buy it for me. And here's access to credit card and like all this stuff and it'll just go out and do it. Like, I don't know if I'd give it that much control, but the like.

TJ (39:32)
Hmm. Yeah.

Chris Gmyr (39:53)
day-to-day automation of things, especially if you're sitting there watching it. It seems a little bit more appealing. And it is on Chromium. So you can have all your same Chrome-based plugins and stuff like that on there too, which is pretty nice. Yeah, don't have much more based on that, but it seems like it's pretty cool and might be tinkering with it more into next week. See how goes.

TJ (40:15)
Yeah, man. That's cool. I look, I have the Claude extension for Chrome. Chrome's not my I use a Zen browser. So it's like I don't have access to like any of this stuff right now. But I do have the Claude extension for Chrome. I just don't ever have use cases where I think it'd be like super useful. Like, I guess, I don't know, like, do I go to my barber's website and be like, yeah, book me an appointment? But like, it takes me like five clicks and I could do that.

by myself too, and it doesn't have access to my calendar. it's like, I don't know. haven't outside of like summarize this blog post I'm reading right now or something like that. Like I don't, I just don't have use cases in mind for it, but it sounds like people do. I just, it's my problem. Like I'm just not creative enough.

Chris Gmyr (40:52)
Yeah.

Yeah.

Yeah, don't know. Yeah, it might be, but I don't know. There's a lot of things that I wouldn't use it for, because I feel like even to get the prompt right to schedule your haircut, like you need to know context about your calendar. So that may or may not be in whatever calendar application you have that's also available via the web. You would have to say like your preferences as well. like aim for a morning, ideally on like Tuesday.

But if that isn't available, like it might be Thursday, like by the time you have to either type or using voice to text, ⁓ now moving forward, you can speak it. The whole process of building that and then checking it to make sure that it actually does the thing that you want it to do, given your preferences that you may or may not put into the prompt or be like that knowledgeable about, like, I don't know.

TJ (41:46)
Right.

Chris Gmyr (42:01)
To your point, it's just easier to click five times, pick the time and date that you want, and just be done with it. It's not that hard, but I see it more for the analytical stuff of people analyzing good versus ⁓ bad performing tweets or Google Ads or Facebook Ads or something like that. I can see it going through and downloading up a bunch of data.

different information with that, like crunching the numbers for you and then giving you like a nice little report after that. I don't do any of that right now, so the automation is kind of less important. And even for Riverside, like I still like listen to things, I still like make sure that it's, you know, tweaked enough that it sounds good. I don't always review the entire show, but I don't know, still need a little bit more like human in the loop instead of just pushing an AI button and hoping for the best.

TJ (42:49)
Yeah, yeah, no, I'd love that but man

Chris Gmyr (42:53)
Yeah, so we'll see. But I know we're getting a little late on time. Do I talk about some Prism updates?

TJ (43:01)
Yeah, sure. I don't know if I have anything like huge on the Prism update front. I've been just like trying to slam through open PRs and issues and comment on things, resolve stuff, close things out. You know, and it's forcing me to get a little bit more focused about, you know, what is the vision for Prism? So like I think I said in the last week,

There was a PR to add like file upload functionality to prism ⁓ And I decided that like I really want to keep prism focused on like generation and decided to like not pursue that PR So trying to just like things like that. I'm just kind of like really getting a little bit more focused about like what prism is And what it's going to be? So the other thing

Chris Gmyr (43:27)
Mm-hmm.

TJ (43:48)
guess like the other like bigger decision that I've made is I think I'm going to punt on 1.0 for a little bit. there are maybe a couple of things that like might be in flux. I'm not a hundred percent sure yet. There's, there's a couple of PRs out there that I think might shift the way I think about things, but,

I'm going to approach the version strategy a little bit different. So right now I've been approaching, like, the version is like semver, right? But like, we're not breaking out of like 0.x right now, right? Because we're just, going to be like floating in like 0.x for a minute. So instead of treating it as like breaking feature fix, we're going to keep it at like 0.0.

And we're going to do breaking and then like fix feature. like, for example, we're like 0.96. So like if I release a new feature, like a non-breaking feature, or if I release a, or a bug fix, we're going to be 0.96.1. But if I do integrate a breaking change, we'll go to like 0.97. So to kind of like just introduce some stability and diversioning for right now.

Chris Gmyr (45:00)
Mm-hmm.

TJ (45:00)
so that people can lock on like 0.96 and still be getting like features and fixes instead of having to be like a little bit more explicit. So I think that'll offer a little bit of stability in interim. And then once we get to 1.0, we can go to like regular semver, think of like breaking feature patch, you know.

Chris Gmyr (45:22)
Yep, yeah, I think that makes a lot of sense. So you don't get up to like 0.385 or something like that.

TJ (45:30)
I mean, I'm

not so worried about that, but just like offering some of like some more stability and like locking your versions and like having an understanding of like what's actually changing in version numbers. Cause right now it's like.

If I go to like 0.97, like you don't know what you're getting. You don't know if it's like a feature or a breaking change now. Like you have no idea what you're getting into. And like I am introducing breaking changes here and there. So like people need to be able to lock their versions in like in a smart way and know when things are breaking and when they're not breaking just by like looking at the version number. So I think this, this change kind of offers that stability, but then yeah, also kind of controls that number from getting a little.

little too wild.

Chris Gmyr (46:14)
Yeah, no, I think that's a really good plan. Because then you can. Be more assured that your whatever minor version that you're locking into like there's not going to be any breaking changes, any issues that come up. It's just going to be like a net new non breaking functionality or like small bug fixes and tweaks that you may or may not actually be using anyways. Yeah, keep an eye out for the larger minor releases.

and then update accordingly from there. So yeah, I think that makes a lot of sense.

TJ (46:45)
Yeah, the other big thing, I just shipped it this morning. There's been...

There's been issues with the new Sonnet 4.5 model and doing structured output. It's something I ran into in my AI thing that I've been working on for Geocodeo. I actually ended up switching over to OpenAI for a little while because you can get very consistent structured output with OpenAI. But with Anthropic, just a challenge because they don't

I don't know. It's not like a first party thing like open AI is. So you end up like sending a message being like, Hey, give me a structured output in this format. Or we give it a like JSON conversion tool and say like, Hey, use the JSON tool to output valid JSON. but it's very inconsistent and like the most inconsistent thing it does is it wraps it in like markdown backticks. and so somebody opened up.

this morning to like sanitize the back ticks out to like hopefully get the valid JSON on the inside of those back ticks to come through. So hopefully that makes it more consistent. Something that like I'm going to end up trying out probably later this week. But yeah, we'll.

We'll have to see like what's next. don't, I don't know what like the next big feature is for Prism. I think it might be async HTTP calls so that you could call in parallel, like a guardrail and a, like an actual, like a user request.

So there's, there's certain things that that kind of unlocks. I guess you can, you can do it now with Laravel's concurrency pieces of functionality, like,

So that makes me think if I should even do it at all. Like I can do async calls using like Guzzle and Laravel's HTTP client. Or I can just like say in the documentation, if you want to make parallel tool calls, just like use Laravel's concurrency feature. And then you don't have to like, we don't have to worry about it at all.

Chris Gmyr (48:43)
Yeah, that seems like a cleaner solution for sure. And then you don't have to support something super custom when it's already kind of handled in the framework.

TJ (48:52)
Yeah, I think the next major thing for me might be going back and enabling, tool calling for structured output. It's been a thing that's like heavily requested and I'd like to offer it. And we did offer it at one point in time, but it really complicates providers doing structured output and tool calling. so that's the reason why I removed it at one point is it was just like, it was

bloating the complexity so much. But it's such a highly requested feature that I think I'm going to have to go back and figure it out. So I don't think that's going to cause any API changes or anything, because that should just be like an internal like, yeah, we can add tool calling now. But I don't know what of like internal refactors that's going to lead to. Like, I think I'm coming to a breaking point where I've got to like maybe think about some architectural strategies here to.

simplify things a little bit. I don't know. We'll see what's next. But yeah, we can probably wrap on that,

Chris Gmyr (49:42)
Yeah, that makes sense. Cool.

Yeah, one quick thing before we wrap. Just wanted to point out that we have a new contact form on the website. So if you go to slightlykeffnita.fm, there is now a ask a question link at the top. And then you can put in your name, email, if you want to share your info on the show. And then if you have a topic or a question, anything that you want to suggest for us to talk about or answer or go through, you can do that now. it's

TJ (50:00)
sick.

Chris Gmyr (50:15)
just an embedded Google form, so it's a lot easier. You don't need to log in or anything else like that. if you want us to talk about something or ask us a question, we'd love for you all to go to the forum and let us know.

TJ (50:27)
Sick, man, that's so cool. Thanks for doing that. Cool. Well, on that note, thank you all so much for listening to the Slightly Caffeinated podcast. Show notes, including all the links from things we've talked about and social channels are down below and also available at slightlycaffeinated.fm. Thank you all so much for listening. We'll catch you next week.