Slightly Caffeinated

Chris Gmyr and TJ Miller share their latest projects, explore AI tools like OpenClaw and Obsidian, and discuss optimizing AI workflows and costs. They delve into personal projects, tech stacks, and future plans for AI development environments.

Ask a question, suggest a topic
Podcast
Chris
TJ
References Mentioned

Creators and Guests

Host
Chris Gmyr
Husband, dad, & grilling aficionado. Loves Laravel & coffee. Staff Engineer @ Rula | TrianglePHP Co-Organizer
Host
TJ Miller
Dreamer ⋅ ADHD advocate ⋅ Laravel astronaut ⋅ Building Prism ⋅ Principal at Geocodio ⋅ Thoughts are mine!

What is Slightly Caffeinated?

Join Chris Gmyr and TJ Miller as they dive into the world of PHP, Laravel, and all things programming, while also sharing insights on family life and other musings.

Chris Gmyr (00:00)
Hey, welcome back to the Slightly Keff needed podcast. I'm Chris Kamir. Hey TJ, so what's new in your world?

TJ Miller (00:03)
I'm TJ Miller.

man, feels like the last two weeks have been such a blur. was starting a new job and like onboarding, which has been super smooth. and I mean, I dove like right in, I got my paperwork done on like Monday, got a tour of like the product on Tuesday and like Wednesday I was creating repos and cutting codes. I dove in.

feet first and just running with it. And it's been great. I get to work with, you know, Laravel and work with Prism like all day long. I couldn't be happier about it. So, it's, it's just been a blur doing that really that's been, you know, making sure I've got my home base covered, but just kind of settling into the routine and. You know, just enjoying, enjoying being in the thick of it. So.

I think things will kind of like settle down over the next week or so. And I know I've been neglecting Iris in Prism a little bit, plan on getting back to that, you know, here is, and I just kind of like settle in a little bit more. So yeah, it's been good, man.

Chris Gmyr (01:11)
Very cool. Glad you're settling in well and yeah, of getting the lay of the land and then you drive. That's awesome.

TJ Miller (01:18)
Yeah, it's been super fun. Like I just merged my first, ⁓ I'm working in an island right now on my own like Laravel code base, like prototyping, like functionality for this feature. And like, I just kind of like merged in the first, like got like V1 of the feature, like the feature prototype, like out and ready. So, yeah, for, for getting that done in the first two weeks, I'm feeling great about it.

Chris Gmyr (01:39)
Nice.

Yeah, yeah, that's great. It's awesome.

TJ Miller (01:47)
Yeah. How about you, man? What's new in your world?

Chris Gmyr (01:49)
well, kind of similar, just diving into the new team, also on my own little island, a team of one dev ex guy, an AI enthusiast. ⁓ But yeah, like trying to continue to like go through the massive list of like ideas and things that I want to do across the board between like dev ex, AI, API stuff, local dev tooling, like it's a lot.

TJ Miller (01:59)
Yeah

Chris Gmyr (02:17)
So trying to figure out what I can constantly ship on every week and then what I need to plan for bigger things or kick off research for. So yeah, just trying to find the balance between the doing and the planning has been a challenge, but also been doing a whole bunch of stuff also. So we have a ton of repos.

TJ Miller (02:39)
Yeah.

Chris Gmyr (02:43)
typically the same tech stack, but there's a couple of one-offs here and there. So basically, had Claude do a bunch of research with the GitHub CLI of go find all of the most active repos, see if they have a Claude MD file, see if they have rules set up, see if they have this and that and a few other things, and basically brought that down into a markdown file. And I think there's 60 or 70 repos that it flagged.

TJ Miller (03:08)
Yeah.

Chris Gmyr (03:12)
and was basically like, make a Epic and Jira, push all these things to the backlog, and mark down what they have to run. If they already have a Cloud MD, just run the rules plug-in, which ⁓ sets up some hooks. There's a rules extract that goes through the Cloud MD file and then goes through any of their patterns, any of their docs, things like that. And basically,

moves some of that information from Claude and Docs to automated rules and then optimizes the leftover Claude information. So it says, like, this is actually a rule now, or this is in a doc that we reference a different way. So let's clean up the Claude MD file. And that's had some huge savings with that. So those are relatively easy that I can just chunk through maybe two or three a day. But it's still going to take me a while.

And then just doing some dev tooling and plugin changes and looking at some research. We used local stack a bunch for mocking and working with AWS services. And they just moved forward with forcing everyone into a paid plan instead of using a free version of it. So I did a bunch of research for like, because we use obviously AWS for everything.

TJ Miller (04:24)
off

Chris Gmyr (04:29)
We use them heavily for local instances and also CI. So figuring out what other options there are. If we were to pay for it, how much is that going to be? All that stuff. So doing a handful of the not so fun investigative and research tasks on top of that. But yeah, going pretty well so far. Getting some good feedback and some more ideas from people that there's just so much to work on.

TJ Miller (04:46)
Yeah.

Yeah. Yeah.

Chris Gmyr (04:55)
Like I feel like I'm gonna make like a GIF or some Slack emoji or something like that of like me with my hands. I'll be like, I'm only one man. And I don't know, see how that goes. But yeah, loving it so far and so far so good.

TJ Miller (05:03)
Yeah.

That's sick, man. I'm so excited to hear that. Like, DevTools and all of that stuff is such a, like, area of passion for me and always has been. So definitely, like, I'm going to, it sounds like I'm going get to work on a little bit of that at Luma too, which is great. but I'm definitely living like vicariously through you, cause this is like, you're working on some really cool stuff, man. So I'm always excited to hear about it and see what you're tinkering with.

Chris Gmyr (05:34)
Thanks, yeah. Be able to share a lot more soon. And then also, we're going to be talking about a whole bunch of little side projects and tangents and things that I've been going on on the personal side, too. So we'll get into that shortly.

TJ Miller (05:48)
Sick. Dude, yeah.

No, I know you've been like cooking and that's it's been really cool to like see you play with this stuff. Yeah, I can't wait to hear more.

Chris Gmyr (05:56)
Sweet. So yeah, I think we covered new check-ins on jobs and changes and stuff like that. Any new, like, coffee or anything exciting on your side?

TJ Miller (06:01)
Yeah.

I picked up a bottle of that espresso you were talking about and I was at the store and I like looked down. I was getting my like regular iced coffee and I like looked down. like, this is the stuff Chris has been talking about. So I picked up a bottle of it too. I picked it up last night. So I haven't cracked into it yet. But yeah, looking forward to that.

Chris Gmyr (06:09)
yeah.

Sweet, I'll have to report back next week, see how you like it.

TJ Miller (06:24)
Yeah, yeah.

My, ⁓ I think, I think my son's going to get into it too, a little bit. He's been like wanting to get it, like he's 12, but he's like, I want to make coffee. I want to have some coffee really bad. I'm like, all right. So I taught him how to use the French press. And so like he made himself a little French press, but then he like grabbed his bottle of Hershey's chocolate and just like dumped a bunch of it in there. So he made himself his like.

Chris Gmyr (06:48)
you

TJ Miller (06:54)
homemade mocha. I was pretty thrilled with his like creativity, but I don't know how good that was going to taste. Just taking coffee and like throwing a bunch of Hershey's chocolate mix into it. Like, I don't know.

Chris Gmyr (06:55)
Nice.

Yep. That's funny,

Yeah, yeah.

but kids, like they'll, I don't know. As long as it's their idea, they'll eat or drink basically whatever.

TJ Miller (07:15)
He was into it, man. Like, he's made it

twice now. He loves it. So, I don't know. Must be all right.

Chris Gmyr (07:21)
Yeah, Yeah, not much for me on the coffee frump. Yeah, just basically doing the same thing. ⁓ Yeah, nothing too exciting. But yeah, wanted to...

TJ Miller (07:23)
Cool dude.

Yeah.

Chris Gmyr (07:34)
coming up. Yeah, I guess we can jump into some OpenClaw and obsidian stuff. Kind of mentioned it a little bit the last time we chatted, but dove into this a lot more, have like a bunch of changes since then. So last time we talked about the problem of using Claude projects in Claude AI chat and not being able to

TJ Miller (07:40)
Yeah, for sure.

Chris Gmyr (07:58)
share ideas like back and forth. Claw doesn't give you an option to do that. There's no API. It was just kind of like little islands of project and content. And there's a lot of things that I wanted to do where it's like, well, this is like a work conversation, but it kind of deals with like dev-x or this other project or maybe like API changes because it's all like kind of related or there's like key information or context in this other project versus this other project. So it got

TJ Miller (08:25)
Mm-hmm.

Chris Gmyr (08:26)
really messy, but I have a lot of the stuff in Obsidian or I could have moved a lot of the stuff into Obsidian. So basically, chatting with Claude and was like, ⁓ like OpenClaw seems to be like a really good option to do this. So basically what I did is set up OpenClaw on a Digital Ocean Droplet just to get it as a separate like entity. Don't have like a Mac mini or I didn't want to put it on my regular

personal computer. I could put it on the NAS, but I wanted it kind of separate because I just wanted somewhere safe to tinker with it basically.

TJ Miller (08:59)
yeah, you can just like that.

The nice thing about doing that, like that's what I have. I've got Iris on a Mac mini. haven't played with open claw yet, but I would probably just do the same. but the nice thing about giving it that isolated environment like that is you really can give it free reign and not worry about it too much.

Chris Gmyr (09:14)
Yeah, yeah, totally. I'm just like, I'm just going to do the recommended like the $24 droplet because it has a little bit more CPU and RAM and whatever. It's fine. Plus it has like a one click like install with DigitalOcean, which is nice. So you still have to go through like the setup and onboarding, but at least like the install was basically done for me within a few minutes. Then I worked with.

TJ Miller (09:29)
Mm-hmm.

Yeah.

Chris Gmyr (09:38)
Claude to say, OK, here's the plan for Obsidian. I basically just want OpenClaude to be one-to-one with Obsidian. And then it helped me build it out. ran all the things that I needed to run on the server, set up the Anthropic API connections, all that stuff, set up Obsidian headless sync. So basically, the main workspace is the vault. So.

remap the workspace to the vault and that gets synced automatically through the Obsidian Headless worker. So any changes that I make locally or like on any one of the laptops or mobile devices or whatever, it syncs the server. Anything made on the server syncs back to all the devices. Perfect. So then...

TJ Miller (10:21)
And that's using the

Obsidian Sync service, right? You're not using the home rolled or whatever other syncing methodologies, right? That's the Obsidian brand sync.

Chris Gmyr (10:24)
Yes.

Yep, Obsidian

brand, it came out like in February, I believe, so not too long ago. So it's official from them, just like an NPM install basic thing. And then you log in, you have a token. You have to be using their sync product. So I'm already using that. Yep.

TJ Miller (10:38)
Mm-hmm.

sick.

Yeah, that's kind of what I was getting at, right? Because

I know I was using iCloud when I was using Obsidian. I know there were some other methods of doing that too. there, thank you.

Chris Gmyr (10:59)
Yeah, you have to use their sync product. then

this plugs into the headless sync, plugs into their sync product. So you have to have both. But I originally tried iCloud when I first set up Obsidian. But that was super slow. It was wonky. It wasn't finding some sync issues and whatever. I switched over to Obsidian sync. And it's been fantastic and super fast.

TJ Miller (11:06)
right. Sick.

It was definitely like clunky for sure

Chris Gmyr (11:27)
ever since. So I would highly recommend if you set it up again, like just opting in to sync, especially if you want the.

TJ Miller (11:33)
And

it's pretty inexpensive, right? I think it's just like a couple bucks.

Chris Gmyr (11:38)
Yeah, it's like four bucks or something like that for like the basic sync, which I think it gives you like, I don't know, one gig or I don't know, something like that. I haven't looked at the information yet, but yeah, I think it's like four bucks a month for annual when you get that amount. It might be a little higher if you go like once a month.

TJ Miller (11:52)
Yeah, yeah,

yeah, it's four bucks a month annual, five bucks a month for monthly billing. then it looks like that's your only option up to like, unless you're going to go with like commercial or like you want beta features or if you're like wanting to use their publishing platform.

Chris Gmyr (12:10)
Gotcha, gotcha. Yeah, I just have the basic one, which has been more than enough. ⁓ Yeah, it's been super solid. And then you can use the headless sync automatically. And then you just run a couple of commands. You log in. You have an encryption token for your vault, which keeps it encrypted on all the different sides of the sync. And then it was like, yeah, I can read all of your notes and everything else. took a few minutes to sync down to the server, but it was good to go.

TJ Miller (12:14)
okay that looks nice.

Chris Gmyr (12:38)
And then what I did was set up Discord to connect to OpenClaw and set up individual channels to go to map directly to like individual directories in the Obsidian Vault. So I can say like my slightly caffeinated channel goes to publishing slash podcasts and it has like all of the

Transcripts I save like all that in there like all the resources everything that basically gets pushed into transistor I've downloaded and have it all there so I can basically see like all of our transcripts for the past like whatever Episode are we on right now 54 so 53 episodes are in there all the trans the raw transcripts you're tagged I'm tagged all that stuff in there so I can basically ask this channel any sort of question that I want

of like, how many times did TJ and I talk about this or that or whatever, or what's the common themes on the arc of the last few months? So I've done some processing with that, which I shared with you. But it's finding some really cool things that we can talk about in the future of, hey, you mentioned this three months ago, but you didn't actually bring it up again. So maybe this is a good content idea type of thing. ⁓ Yeah. And I went through and now have about like,

TJ Miller (13:41)
Yeah.

I love this.

Chris Gmyr (13:56)
15 or 20 different channels for personal stuff, work stuff, or side projects, like I said, for content and publishing. And then as I was going through this, I'm telling OpenClaw via Discord or even locally with Obsidian of, hey, we just did x, y, and z. Let's do a draft blog post of what we set up and how this all works. So I've also been

like drafting out all this new content to share on the new website of like all this open-claw and obsidian process and all the work and skills and things like that that I've been working on. So it's been really cool to like be able to just kind of pull that trigger of like, hey, this is a really good content idea. And then I actually have a skill for like slash content idea that pulls in like all the contacts or like gives instructions through the input of

like guiding it what to do. And it asks like, do you want this to be a podcast backlog or a newsletter or all of them or whatever? And it'll go through those like, I've set up backlogs for all these different segments. So there's like a social backlog, a podcast backlog, a blog backlog. And it basically maps like all that stuff around of like, like.

TJ Miller (14:53)
Yeah.

Chris Gmyr (15:14)
we talked about XYZ, that would be a great newsletter intro. So it'll add it to the newsletter backlog. So when I am looking for something to write about in the next newsletter, then I can look at that and then be like, OK, this is a great idea. Let's expand on this more. Here's what actually happened after the fact. And then I can flesh it out as I see fit. So it's been really sweet to have this full flywheel lifecycle of doing things to content, to revising the content, to actually publishing it.

getting feedback on that, and then just generating more content from there. And all kind of automatically through the system. And because it syncs everywhere, if I need really heavy processing or just want to work locally, now I can do the same process in Discord locally through Cloud Code, directly connected to the Obsidian Vault, which has also been super cool.

TJ Miller (16:08)
Yeah, man, that's so sick. I mean, one of the things that I've always like, believed with AI is

cooler stuff that you like. I guess to phrase it another way,

If you want to do cool stuff, it takes data, right? Like the more data you have, I think the cooler stuff you can do. like, I know you, we've talked about it early episodes and like we've talked about it forever about different note taking techniques and like PKMs and all this kind of stuff. And, we both kind of like at one point in time, like settled on Obsidian and like really got into that. And like, I tried to get into it.

Chris Gmyr (16:38)
Mm-hmm.

TJ Miller (16:46)
My ADHD makes that like really hard, even though it's like crazy beneficial for the ADHD, right? It's this conundrum. But like, I always imagined like, I would, I always imagined like a cool system like, like that, where it just kind of has all the data from like everything just available at its fingertips. But I'm not diligent enough to like, curate and create all of that data. So it's kind of tough.

Chris Gmyr (16:52)
Mm-hmm.

TJ Miller (17:14)
on my end, but it's so cool to see with like, you, I have seen you be like relatively diligent about like putting stuff in there and keeping it organized and God that makes like such the difference when you turn AI loose on it. Like you can do some really, really cool stuff, but it takes having, you know, that data available.

Chris Gmyr (17:34)
Yeah, yeah, totally. it's been, depending on the timeframe, it's been either easier or harder to keep up with that. I feel like it's been, like we talked about this a little bit too more recently, it's been hard to keep up with all the AI stuff coming out of all the channels altogether. So I feel like I haven't had the time or the energy to read as many blog posts or.

watch videos on YouTube from creators and tech people that I like. So it's basically mostly been all AI all the time. But trying to spread that out a little bit now that I have a better handle on at least the system that I'm building. And it's definitely beneficial. And I'm so glad that I kept with Obsidian because it's just all Markdown files. And now they have more AI tools.

TJ Miller (18:18)
Yeah.

Chris Gmyr (18:19)
Like, it's been fantastic to just, I don't know, be lucky, I guess. The ecosystem that I picked is just well-versed for this. And being able to build in so many, I think I have probably 20 or 30 skills doing different things in the vault that is also passed around to all these different instances. So I can do basically the same thing across the board. And it was really nice in Discord to have that.

clean separation, but the option to go cross boundaries to a different project, those are tied to specifically to the config for the channels in OpenClaw. So I'm like, OK, how do I do this locally? Because I want that kind of cleanish separation and context and doing a handoff of last session, what do we talk about, or what do we want to

make a concrete artifact. So I basically brought all that into the local vaults as well. So I have a context switcher skill that will automatically have a little .file, almost like lock file of, hey, this is the project that we're currently talking about, and a hook that brings in, be like, OK, let's talk about something that I want to publish, or let's work on this publishing flywheel.

TJ Miller (19:19)
Yeah.

Chris Gmyr (19:38)
or some content ideas. So it'll bring in all the content projects, the last context, the last session notes, things like that, and any important skills that are relevant to that context or project, all available in the new context window. So I can say, CTX, podcast. And then it'll do the last context, the last session, bring in some podcasts, like import and publishing skills.

And suggest like where we left off like the last time I was talking about it, which is like super cool and then when I like And then I like do the thing talk about whatever I can also be like this would be like a great blog or something like that and then it'll bring in the context for like the blog project and then I can clear it and switch to a completely different track or ⁓

TJ Miller (20:09)
That is so cool. Like I'm just sitting here drooling. Like that's so dope.

Chris Gmyr (20:31)
like thought process for something about work or some other personal project or something like that. So I've like tried to build in the constraints and the updates that I wanted to do. Because in Discord, it was like tracking all that stuff automatically based on the channel instructions, but losing that locally was kind of a bummer. So now the systems work similar enough and now we have these like tracking files that I can then move from like a local Obsidian and Cloud Code instance

right over to OpenClaw in Discord, if I'm mobile or something like that, to basically continue the conversation because it knows what we were just talking about and the actions that we just took. So that's been super helpful to just pick up from the computer and just go wherever. Start making dinner, check in on this, or have it read something to me, or who knows what. It's just been like,

TJ Miller (21:17)
Yeah.

Chris Gmyr (21:25)
Pretty awesome to try and make the two types of systems function similar enough that I can use it autonomously no matter where I am or what device I'm using.

TJ Miller (21:31)
Mm-hmm.

Dang, dude.

Dang. That is so sick.

Chris Gmyr (21:40)
Yeah, it's been pretty awesome. Some downfalls of the system, if we wanted to get into that, is originally, when I set it up, OpenClaw, up first, I was just using the Anthropic API connection to use Sonnet and Opus. The default was Sonnet. But between all the setup and...

TJ Miller (21:46)
Yeah, let's do it.

Chris Gmyr (22:01)
churning through all these different iterations and trying to get everything working really well. It was going through like $20 to $40 a day to just chunk through this and get it all set up. like, this is not sustainable. I will not be able to use the system at all. So after I got it to where everything was functioning and basically working OK, I'm like, OK, what are my options for

TJ Miller (22:11)
Mm-hmm. Mm-hmm.

Chris Gmyr (22:29)
other services, other models to kind of get a similar output to Sonnet, but not like completely break the bank if I'm using this consistently like day to day. So I came up with like a short list of like the Gemini Flash 3 preview or something like that was really good and like fractions of a penny compared to Sonnet. And then doing a couple little things with like GPT like

mini something or other, and then a fallback to Sonnet. So if any one of those gets rate limited or whatever, OpenClaw will switch to the next fallback model and keep on going down like that. But I think I have the default of just the Gemini 3 preview, which has been pretty solid. And it's a lot cheaper than Sonnet for daily use.

TJ Miller (23:08)
Yeah.

Chris Gmyr (23:22)
⁓ So I've definitely like moved a little bit back towards, cloud code locally with the obsidian wall and just like doing my main kind of work with that. Cause typically like during the day, like I am on a computer, so it's not that big of a deal. So it's just like, okay, let's not use open claw and discord for everything. Cause I was, you know, just trying to test out the system and see how everything functions and getting it kind of set up how I wanted.

TJ Miller (23:35)
Great.

Chris Gmyr (23:48)
But now that I pulled some that functionality back down locally, I'm able to do like more and bigger processing loads with my subscription locally and not having to worry about like API pricing, which has been really nice. And then switching to the Gemini model in OpenClaw as the default, that's been able to save like a whole bunch of money as well too. Because it's like dollars a day instead of like...

TJ Miller (24:11)
Dude.

Chris Gmyr (24:15)
20 to 40 a day. So it's been a lot.

TJ Miller (24:17)
Yeah, yeah.

When I get into like high usage with Iris, like I'll blow through 25 bucks a day. Like, it's scary. However, it's been really cool. Someone in the Prism Discord that bought Iris, I added multi-provider support to Iris. So you could always configure the models, but I...

didn't have it so you could also configure the provider so you were like stuck with Anthropic. But I added multi-provider support and they were playing with OpenRouter and they got their cost down to like a dollar a day and it's been like pretty effective for them. So yeah, there's definitely like tuning that cost down is takes a little bit of effort, but you can absolutely do it.

Chris Gmyr (25:05)
Yeah.

OpenRouter was something that I was going to look into too, but I just wanted to get the system set up and running. But now it probably makes more sense to go the OpenRouter route, because now I have some other things running in the background. This content flywheel type of thing is a cron that runs on Sundays that does a bunch of work. And then I have a daily like

TJ Miller (25:07)
Yeah.

Chris Gmyr (25:31)
podcast like fetching service that so I listened to a bunch of podcasts a lot of those don't have like transcripts or notes or anything like that so I can basically like dump a URL in here and be like download this episode or like track this like season of this podcast and it'll like look at the you know Apple or Spotify feed it'll pick up if it does have a transcript available it'll use that if not it'll take the audio

It'll push it through OpenAPI. So I set up API credentials and Google credentials and a couple other things. It'll transcribe the podcast, put it in a note in Obsidian. It'll process it. So it's all set up in the note. And then I can further process that. So I have a couple of blog posts out right now and a link to the show notes. We're recording on Wednesday. I have another one coming out tomorrow.

TJ Miller (26:07)
Mmm.

Chris Gmyr (26:24)
for the podcast transcription service and flow that I built. ⁓ So that'll be all available. But yeah, it's just like, how can I make these things better in the least amount of work to actually trigger the thing? And since I listened to so many podcasts and it's like, this little thing relates to this other thing over here. And especially in the personal or financial side, relationships.

TJ Miller (26:28)
sick.

Chris Gmyr (26:50)
and especially tech, it's like, these three different podcast episodes map to something that I'm really interested in. So now I can download those automatically, have the shown transcription generated automatically, have key takeaways, things like that, questions that were answered, and say, make me a note that encompasses these three episodes, and having some action items to look into or research or do or.

whatever, that it would have been such a manual process to do that before. I don't typically do that. I might make a note if it's really important or throw something into Todoist, but typically I would just forget about it. And maybe the next time that little idea came up, I'd be like, yeah, that was a good idea. I should do something with it. And now I can just throw it into a workflow and have it pop out a note. And then I can take action on the note or put something into Todoist or connect those dots right there.

TJ Miller (27:28)
Yeah.

Chris Gmyr (27:46)
and when I'm actually still thinking about it, which has been pretty sweet.

TJ Miller (27:50)
Yeah, dude. Dang. Like, this is sick. I might have to play with this. I've always kind of envisioned a like... Like, with Iris, we have the memory system, we have the truth system, like, those are all cool, and like, they have their purpose, but I was thinking for...

like collaborative notes or like longer term memory and concept connection and let ability to have like a different medium as shared space, like this, this world of notes, right? So it was kind of envisioned like this notes feature for Iris. And I think maybe this is like the route to go is like put it into obsidian. That way you kind of have this like really polished experience for the note taking side of things. But then you have.

this agent that's aware of all of this stuff too, which I don't know, man, that's pretty sweet. The other thing that could be interesting too is giving OpenClaw access to Cloud code and put Cloud code on the DigitalOcean droplet. And then it can defer to that as well. And that might help with some of the split between using your Cloud plan versus your API.

Chris Gmyr (29:05)
Mm-hmm. Yep.

TJ Miller (29:06)
kind

of let it like defer some workloads off to cloud code.

Chris Gmyr (29:09)
Yeah, definitely. Yeah, could be an option. Yeah, just doing what I can to tinker with it and not blow out my budget for sure.

TJ Miller (29:15)
yeah.

Yeah,

definitely. like, yeah, yep. Yeah, I built a, it worked part of what I built. I like, I built the feature and then I built a LLM as judge, like evaluation system to like evaluate the outcomes of the process.

Super valuable, but like, yeah, Easter tokens.

Chris Gmyr (29:39)
Yeah, yeah.

was in the new updated skills creator in Cloud Code. It comes with a whole bunch of eval functions and workflows and things like that, which is super cool. Yep. So you do slash skills creator eval and then your skill and then a description.

TJ Miller (29:54)
really? I haven't seen that yet.

Chris Gmyr (30:05)
⁓ It'll go through and figure out what the skill does and then come up with this like whole matrix of using the skill versus not using the skill. It'll bring up like a little UI so it'll like do all the work. You have the prompts and you give it feedback on like how it's doing versus like the skill or not skill usage. So for like three different prompts, you'll have the six inputs that you can give feedback on.

TJ Miller (30:18)
Mm-hmm.

Chris Gmyr (30:31)
So the three prompts versus skill or not skill. And then you submit that. It'll download a little JSON. You transfer that over to Cloud Code. Here's the evals. It'll go through and rescan everything and adjust the skill. It'll give you a new set of evals if you want that. So I did that with one of my skills yesterday. And it flew through my Cloud Code session for the next couple hours because I went through it four different times.

It's like, damn, that's a lot. the output was very good. So it probably depends on what the skill is. Mine was a lot of this crunching of the notes. And there was a ton of notes and transcripts and things like that. So it was a lot of data that I was trying to crunch through as well. So I'm sure a simpler skill would not eat through that much session usage.

TJ Miller (31:01)
Oof. Yeah.

Yeah.

Chris Gmyr (31:26)
Just depends on what the skill is actually doing. But yeah, there's a ton of goodies in the skill creator skill that they pushed out probably a month ago or something like that. I think they had the 2.0.

TJ Miller (31:39)
Yeah, I don't know if I've messed with it since the 2.0. I know all of my skills inside of Iris and what I'm using at work and everything. I know they're like kind of stale. I feel like I was really ahead of the curve for a long time with like all this AI stuff. And I feel like everyone's kind of like caught up.

And so I'm at this place where I'm like, all right, now I need to like go back on all of the stuff and like refine it as things have continued to evolve more. cause I kind of like stagnated while everyone caught up, you know, and I I'm due to kind of like revisit all of that stuff. So it's really good to know because I've got like, yeah. Four or five like core skills that I would love to get, you know, more feedback on from something. So if they've got like evaluation systems, I'll probably end up pumping them through that too.

Chris Gmyr (32:23)
Yep, it is a full eval system and then you can just say what can be improved upon. And it depends on the skill if you want it to be auto run or if you're manually running it too. It's good to push that through because a lot of the skills that I have are manual. So I don't want you to infer anything from the conversation. I want to run it as a command because commands are now baked into skills. ⁓

TJ Miller (32:49)
Skills and

commands got merged into just skills. there's like a property, my gosh, the word just disappeared from my head. Front matter. There's like a front matter key for whether it's invocable as a command or not. I think it's like for agent, or I can't remember what it is. But there is a flag in there that will allow it to surface as a command or not surface as a command.

Chris Gmyr (32:51)
Yeah.

Mm-hmm. Yep. So things like that have been really good. You can also put front matter in to specify, like if you have an agent to do certain work, like you can add that to the front matter for a skill. So when you call this skill, I only want you to use like this agent. So it could be good for, like if you, like I know you have like a Laravel and like PHP skill. If you have like ⁓ a PHP like agent, like you could specify that and it's not like in the main agent.

window that will do that. Also with Opus, can select low, medium, high effort as well, because that's in the new 4.6. So there's so many new things that you can do with skills. And the skill creator will help you go through the process of revamping some of those, because I've done that with some of my older skills and some things that work in the plugin system. I'm like, hey, let's check out this skill and revamp it.

TJ Miller (33:42)
Mm-hmm.

Chris Gmyr (34:02)
make it a little bit better. yeah, Skill Creator is like a huge, huge tool.

TJ Miller (34:07)
Yeah, yep. I'm going to be helping at Luma get AI development environments. I'm going to end up evangelizing all of that stuff and getting it configured and helping set the tone and standards and everything with all of that stuff. So I'm due to take a good pass at it. So yeah, I'll check out the Skill Creator evals. Because yeah, the Laravel and PHP ones, they're super effective.

they're due to be better.

Chris Gmyr (34:36)
Yep.

TJ Miller (34:38)
Sick man. Well, on that note, do want to wrap up?

Chris Gmyr (34:41)
Thanks for listening to the Slightly Caffeinated podcast. Show notes and all the links and social channels are down below and also available at slightlycaffeinated.fm. If you have any questions for us or content suggestions, go to the Ask a Question page on our site and we'll feature it in an upcoming episode. Thank you all for listening and we'll catch you all next week.