Slightly Caffeinated

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

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

man, I have been just super family oriented, household project oriented, like crazy the last two weeks. Definitely been pushing on Iris and doing some cool stuff there.

But my son got super into rollerblading and so I've been like taking him to the skate park and like I've been jamming onto my skateboard there. I've got some aggressive inline skates coming soon. we were talking a little bit before the show, like I have crazy wide feet and so it's impossible to find skates that fit. But I have a lead on some. They're coming, they're on back order.

My wife is actually graduating with her associates tomorrow. We're recording on Thursday. So yeah, her commencement's tomorrow. And her dad's coming to town for the commencement. So it has been a full court push on like wrapping up house projects and deep cleaning and like trying to get some yard work done so it doesn't look like I'm

Chris Gmyr (00:53)
sweet.

TJ Miller (01:14)
a terrible adult. So this involved like, I started, I think on Friday, I like, I painted our whole living room. And yeah, it was a couple late nights. And my back is just wrecked from rolling them out. But yeah, we had like,

Couple with patches and all the walls had patches and like the paint was all scuffed up and horrible because we did a I Figured a matte finish would be cool. Never never again. Will it ever paint a wall with a matte finish? It shows every fingerprint ever all over the place so yeah, I repainted all of that and then Just my gosh, yeah very

very domestic focused with just everything and jamming along at work doing some really cool stuff there too. So yeah man, how about you?

Chris Gmyr (02:09)
Yeah, similar. I've been busy doing a bunch of traveling. after we chatted last time, had a scout camping trip to go to was super fun. It was about like an hour and a half north of us. It was on a lake. So the views were great. The kids went fishing Saturday and Sunday mornings and went on like a little hike and did like a Junior Ranger program with the Ranger out there at the park.

I learned about a bunch of local wildlife and plants and all that that the kids had fun. Ended up raining a little bit on Saturday night. So a little bit of a bummer there because everything gets wet and then it just takes so long to clean everything up and get it all dried out and repacked for next time.

TJ Miller (02:55)
man, having

camped in the rain in a tent before, the like, the unpacking the tent up and having to unpack it and dry it all out and just, it's a nightmare.

Chris Gmyr (03:06)
Mm-hmm.

Yeah, yeah. Luckily with the truck and like I have a bed liner in there. I don't even worry about it. I just kind of do my best to fold it up a little bit, but I don't pack it into the bag or do anything with the fly. I just kind of like fold it in and throw it in the back of the truck and it's like, OK, I'll deal with this when I get home. So it's not like you're throwing it in your backseat or anything like that. But yeah, and got back on.

TJ Miller (03:20)
Hmm

Yep. nice. Yeah.

Chris Gmyr (03:34)
that Sunday, then got ready to go to our engineering offsite in Chicago. So I left for that on Monday morning of last week. It was a great time out there seeing everyone in person, seeing some new faces, seeing some old faces, seeing a bunch of the old Curalogy people also that I work with, which is pretty awesome.

TJ Miller (03:52)
Wild.

Chris Gmyr (03:53)
Had a good time, had like a hackathon. So if you have time, we can talk about that later. And yeah, just super nice to see all the new people on the org and the team and stuff like that. Then came back last Thursday. And then my mom also came in on that Thursday. We went to the beach. So on Friday to Sunday, because it was my daughter's birthday yesterday.

TJ Miller (03:59)
Yeah.

Chris Gmyr (04:21)
And so she wanted to go to the beach and hang out there. my mom came down. We went to the beach and hung out out there and did a bunch of beachy things. rained a little bit out there, but we still made the best of it ⁓ and had a good time. And yeah, just now trying to recalibrate back at work. And yeah, I feel like I have a lot of stuff to catch up on and get going with. But yeah, it's been a couple of busy weeks. We had to do a lot of prep.

TJ Miller (04:31)
Yeah.

Chris Gmyr (04:47)
You know, like you were saying of like being out of town and family coming in and other events and stuff like that. but got through it and, ⁓ relatively unscathed. So yeah, not too bad, but yeah, just happy to just have a couple of quieter weeks right now. ⁓ to just kind of get back into things and keep on plugging away at the regular house projects, not the extra house projects and, ⁓ regular work routines and all that stuff with the kids and whatnot.

TJ Miller (04:56)
Nice.

Yep.

Chris Gmyr (05:16)
before school is out sooner rather than later.

TJ Miller (05:19)
my gosh.

Yeah, my son, my son said something to me the other day. He's like, yeah, we only have like a month of school left. What happened? Why? Where? Where did this come from? Yeah, like we just started. Yeah. Feeling out.

Chris Gmyr (05:30)
Like, oh no, I thought we just did this.

Yeah,

it's I mean, lots of that's good things, but just, you know, out of the ordinary. So looking forward to getting back to the routine and all that. ⁓ So I thought I would have some good. Coffee stuff to talk about, but unfortunately, it was not very good coffee throughout all of the adventures. So we talked about Camp Coffee before it is what it is, little chunks in it keeps you going.

TJ Miller (05:45)
Mm-hmm.

⁓

Chris Gmyr (06:01)
⁓ Chicago was okay, but didn't get a chance to go to any actual coffee shops. So, it was just like hotel coffee. I think they had some Starbucks in there too. cause in the place that we stayed, it was the hotel and like conference, area, like in the basement. it was, basically all catered the same way. so just the same coffee every day and mostly like not like room.

TJ Miller (06:21)
Yep.

Chris Gmyr (06:24)
hotel room coffee, but like, you know, the catered, you know, breakfast coffee type bar and just drip coffees and stuff like that. And then beach coffee and like one decent one. But it was fine. It wasn't like the greatest coffee shop. So kind of, you know, all all Mrs. On the good coffee train. Best coffee is Monday morning when we made it ourselves here.

TJ Miller (06:28)
Mm-hmm.

Yeah ⁓

Chris Gmyr (06:50)
Yeah

TJ Miller (06:50)
Yeah, yeah, so I don't know if I talked about it last time, but like my wife finally talked me into going through the Costco, the Costco coffee she bought. So she like, I don't know, like a month or two ago, she just decided to order a giant bag of like Costco espresso blend. And I put that in our like dry food storage. I was just like, yeah, we'll just store this one and like drink other coffee.

but she's like, you know, we really need to break into that. I'm like, all right, fine. So I did. Not so bad. I was really dreading it. I thought it was going to be terrible. I was really not looking forward to it, but it is so much smoother than I thought it was going to be. it's actually been like a really pleasant cup of coffee. So yeah, all that, all that dread and putting it off to be pleasantly surprised.

Chris Gmyr (07:42)
Yeah. So you're going to have to drink the rest of it, or are you still saving it in storage for portions of it? Or what's the plan? Just finished off?

TJ Miller (07:52)
Now we're just

we're gonna go through it. So the bags open, we'll end up finishing the bag. I don't know when that's gonna be because this is a giant bag of coffee. I think it's like two pounds or something. I mean, it's Costco. So, you know, it's huge. yeah, we're just gonna trudge through it. But it's been it's been nice. I don't know what I'm gonna do afterwards. I kind of want to mix things up. Because we've been doing the same thing for the last like several months. So we'll

Chris Gmyr (08:01)
it

Yeah.

Yeah.

TJ Miller (08:20)
We'll see what we move over to, but yeah, pleasantly surprised with the Costco whole bean.

Chris Gmyr (08:27)
Nice. That's cool. Glad it worked out. Yeah. I was really thinking after a couple of weeks off, you'd go get your Turkish coffee.

TJ Miller (08:30)
Still no Turkish yet. Still haven't made it out there.

I know, I

know, uh, where I mentioned it to my wife, so now it might actually happen. We'll see. Yeah. Yeah. Yeah. They had like, uh, really good food there too. It's like a little cafe, like a little, yeah, a little cafe. So like, I think, I think we can make a thing of it and like have nice like lunch dates or something. Um, yeah.

Chris Gmyr (08:46)
Yeah, nice. keep us posted. Hope you get it.

Yeah,

do it. That's your homework for next week. Come back and report. See how your lunch date in Turkish coffee went. Cool. Sweet. So yeah, next up we have, yeah, we wanted to talk about the new cloud usage limits.

TJ Miller (09:06)
It's my homework. Yeah. Yeah, I'm to do it. I'm going do it.

I mean,

let's take a step back and talk about just cloud usage as a whole. I'm on the 5X plan, like the $100 a month plan. And I, for the longest time, like, I mean, a fairly heavy user, but definitely not having multiple projects all running at the same time.

Chris Gmyr (09:33)
Yep, same.

TJ Miller (09:46)
kind of scenario, but after they put in like peak hours, man, I was hitting limits left and right, like multiple times a day. I was hitting my like daily limits and it was so frustrating because I'd get into a position where, you know, we talked a little bit about Gent, David Hemphill's project and I would just drop in a PRD

Chris Gmyr (10:05)
.

TJ Miller (10:11)
and just breaking it up into tasks was taking up 50 % of my daily budget. And I'm like, we're not even doing any work yet. We're just getting things situated. Normally, I would have gotten through most of the day of heavy usage without even hitting them. And then, yeah, we're just now hitting them in the planning stage. Brutal.

And that's even still optimizing between like PRD generation being Opus and then switching to Sonnet for any work after that. So it's like, I'm just using Opus for planning. But yeah, it'd be like, yeah, 50 % of my like daily butt or not daily or like whatever out, however they chunk it up, right? It's like every like few hours, right? So I was like hitting my like those limits like constantly.

Chris Gmyr (10:57)
Yeah.

TJ Miller (11:04)
Never hit my weekly limit, but was hitting those daily ones like two or three times a day. And it was so annoying. I couldn't believe it. And for the longest time, if you like look at the clod code subreddit, it's always people complaining about, ⁓ dude, like degradation or like this and that hitting limits. And I was always like, what are you guys doing where you're hitting these limits constantly? Like, are you really?

like Tony starking it with like five projects all going at once 24 hours a day like I don't know so I was just like I never understood it but yeah after they put in like peak hours it's been terrible experience it was getting to the point where it was getting hard for me to be a fanboy like this is this is actually painful and I was

deep in some iris work and almost went and like got a Codex subscription because I'm like this is outrageous. So what was it yesterday? Yesterday they like basically came out and said we have a new partnership with SpaceX which I don't know how I feel about but was also like interesting.

Chris Gmyr (12:04)
Mm-hmm.

TJ Miller (12:15)
you know, they've got a new partnership with SpaceX and along with that and some other partnerships for compute power, they are broadening the daily limits, I think, and getting rid of peak hours.

Chris Gmyr (12:28)
Yeah, that was my understanding too. I haven't looked totally into it yet.

TJ Miller (12:34)
I know

for sure they said that they were getting rid of peak hours. I think they also said that they were expanding the daily limits.

And that was for Cloud Code accounts and API accounts as well, like APIs as well. that was like kind of spanned like both those products. So the API side and Cloud Code.

Chris Gmyr (12:56)
Yep. Yeah. Nice. That'll be good. cause I've ran into the limits a little bit, like one time recently with a gent, like you were explaining, like with the planning PRD breaking up individual tasks. then the one time before that was when I was doing like a lot of, data crunching over the podcast, transcripts and notes and things like that. And doing some like evals on

some of those skills. And it just took a lot of content and going back and forth. And that was the only other time that I've reached the limits. But I'm split. I only use the 5x Max plan on my personal computer for my own things or side projects or whatever. So at work, I was on the console API plan.

TJ Miller (13:31)
Mm-hmm.

Chris Gmyr (13:47)
But now switched over to our usual AWS Bedrock plan. don't have to worry about any usage limits over there right now. They will be kind of more self-imposed in the future once we kind of get a good level on everything and usage for everyone. Because we'll probably put like, you can't go past like, don't know, 100 bucks in credits or whatever it costs.

is associated with bedrock.

TJ Miller (14:10)
Yeah, so you're

using Cloud Code via the API then, right?

Yeah. Yeah, we,

Chris Gmyr (14:16)
So a little bit different and Bedrock

is a totally different setup.

TJ Miller (14:21)
Yeah, that's interesting. were using Eluma. So I have a 5x plan personally. And then I also have API credits and stuff too, because I use Iris a lot. And that's all through the API. But then at work, we have just the enterprise plan. And that's what we're onboarding everybody into, is the clogged code enterprise with the like.

When you have like an enterprise seat, think there's two levels. have like, like some increased limit one and then like your regular seat. And so like all the devs are going to be on the increased, the increased, like the premium seat. I think that's what it might be. don't know. So they have like a regular seat and a premium seat with like more token allotments. So we're doing that and yeah, we'll see how that works out. It's gotta be nice not having like technically any limit though. That's pretty cool.

Chris Gmyr (15:07)
this.

Yeah,

yeah, totally. So it's been nice. But yeah, I'm glad that all the usage limits are getting fixed because they were a lot better. And then it kind of seemed like it went downhill for a couple of months there with the increased limits and the different hours that they were doing the different limits. So hopefully they'll just kind of ride the ship and keep everyone happy because it is pretty annoying when you get close to it.

Like, well, I can't even start the next task because I 10 % left and two hours to wait type of thing. So yeah, it'd be good to see how that works out. yeah, glad there are changes in rate limits.

TJ Miller (15:43)
Yep.

Yeah, this reminded me of, this reminded me, I've been sending you a bunch of reels on Instagram and this dude, I don't remember his name, we'll have to link it in the show notes, but there's this dude who like records these reels talking about AI and like AI development and like Claude, Claude code, codex, all that stuff. And his reels are him.

Chris Gmyr (16:00)
Mm-hmm.

TJ Miller (16:19)
He's got some that are just like talking head reels, but like most of them and the ones that I really enjoy are him just like cruising around a neighborhood on a one wheel, like an electric one wheel. And he's just like cruising around dropping straight facts. Like I, I've not seen him make a point that I haven't agreed with yet. So it's been like really refreshing to also get like a lot of validation on the way that I think about things and the way that I work.

⁓ but he also has like really interesting, I think solid takes on like the industry and where like the industry's at, where it's heading. and I'm just so amused by it. So like, talking about like the cloud usage limits just made me re I just watched a reel of him talking about it. so we'll have to link in the show notes cause it's just great. Like I, I want, I so badly want to get a one wheel now and just like ride around talking about code, you know?

Chris Gmyr (17:14)
Yeah, it does look like fun. His name is Ed Honor. So I'll put that in the show notes as well. But yeah, they have been pretty entertaining. And I'm like solid information, like you said. So I'm like, yeah, I'm like not long. I'm like, oh, this is cool.

TJ Miller (17:24)
Yeah, it's...

like the last place I expected to be like picking up my tech news is from Instagram Reels, but it's, ⁓ he dropped great stuff. And then, yeah, I wouldn't have known about Claude rule files, like the rules. And I saw that one in Instagram Reel too. So like, yeah, definitely not the place I expected to get, you know.

that kind of info, but it's working.

Chris Gmyr (17:53)
Yeah, yeah, I feel like between like Instagram and Twitter, there's been some like good stuff, but you have to get it like early. And it's hard to tell like what's early on or not, because after, you know, a few days or a few weeks, even like everything is just like regurgitated like slightly. And it's just like, well, is this even relevant anymore? Is it just someone who's like jumping trying to jump on the bandwagon of this new?

TJ Miller (18:07)
Mm-hmm.

Yep.

Chris Gmyr (18:21)
package or plugin or different, you know, workflow to do certain things and it's really hard to keep up with all of it and actually like validate it and I've put a couple that I've like found either like plugins or repos or something like that of like, like this is the next, you know, best thing for, you know, X, Y, and Z. And then I give it a Claude and I say like investigate this repo.

TJ Miller (18:29)
Yeah.

Chris Gmyr (18:47)
see how solid it is, see if there's any security implications, things like that, and do a much deeper dive of should I actually use this as a valuable, or is it going to open up a whole can of worms or security issues or whatever. And the vast majority of them are like, nah, you probably shouldn't use this. It's like, this is the package, here's what it does.

TJ Miller (19:05)
Yeah.

Chris Gmyr (19:09)
It has like these 10 dependencies and dependencies of dependencies. And this is really sketchy over here and that I wouldn't trust. So it's like, yeah, you probably shouldn't use this. Or if you want to try it on a personal project, but you shouldn't typically use it for like a work project type of thing. So it's like, OK, well, can't do any of that. So I'm very picky on what I actually try and like introduce for tooling.

TJ Miller (19:25)
Yeah.

Yeah, if you really sift through the weeds, regarding the muddiness and the regurgitation of all of this information and how that happens, it's this weird game of telephone. If you spend a little bit of time digging through the weeds, you will find people who are the source disseminators of that.

And that's kind of who you follow. takes an eye. takes taking a step back and thinking about the content you're consuming and where you're getting it from. there are definitely people who are source disseminators of this stuff. And if you can find them, kind of like this dude on Instagram, it's really valuable to like, those are the signals through the noise.

Chris Gmyr (20:19)
Yeah, yeah, totally, 100%. And I've started a private Twitter list of trusted AI accounts. So it's a lot of the anthropic stuff. All the anthropic and cloud code devs are on there. People like Matt Stauffer, a couple of the AI people that I follow, things like that.

TJ Miller (20:27)
sick.

Chris Gmyr (20:42)
I think it's just a private list, but your flavor is going to depend on what you like or who you want to follow and things like that. yeah, that's one the things that I check. I try and do it on a daily basis. So just going through there and be like, what's new? What's the latest from Anthropic? And it kind of ties in well with the Releasebot stuff and some of the blogs and videos and YouTube things that I follow as well. But yeah, it's just a lot of content to keep up with and sift through.

on a daily basis.

TJ Miller (21:10)
Yeah, I've been so overwhelmed. I've like really unplugged from a lot of things. Like I haven't, I haven't checked for my Twitter and it's been a minute. It's been a few weeks since I've even opened it up. I've just had to take a step back from it all and, kind of reevaluate and reprioritize things. So like I've been so out of the loop on stuff that like I'll, I'll scroll Instagram reels, but like

I know, man, I haven't been back to Brave Twitter yet. I think I'm building back up to it in the next week or so, but yeah, it was too much noise in my brain, man.

Chris Gmyr (21:47)
Yeah, yeah, 100%. I've taken a little break the last couple of weeks too with all the travel and activities and extras that I've had to do around here. hopefully like later this week into the weekend, I can kind of get back into my routine with that of like, you know, content and just grabbing new information, tinkering with stuff. But yeah, it takes a lot of mental overhead. So it's been kind of nice little break.

TJ Miller (22:03)
Yeah.

Chris Gmyr (22:14)
You know, of what we've talked about, like in the past to just like just mentally just take a break, focus on the things that you actually need to focus on and then reset a little bit. So that's been nice.

TJ Miller (22:23)
Yeah, for sure, man. So kind of moving forward, we have some skills we want to talk about. We've been talking about, the last few episodes, skills we're using and working on. I've talked about my PRD generation skill, my code review skill. I know that you've talked about some of your

⁓ like wiki ingestion skills, some of your like other work ones. And so I think, I think these are a couple new ones that we, you may have mentioned that we haven't talked about yet. And I'm really interested in hearing a little bit more about them because I think. I think it could be relevant to some stuff that I'm starting to work on. So, would love to hear more about it.

Chris Gmyr (23:03)
Yeah,

totally. So the three skills that I've been actively working on at work are all about session and context information. So the first one is the simplest where it's just a session brief. So this is kind of similar to compact, but it gives you a little bit more

control and information in there. So Compact is just built into Cloud Code and it kind of guesses what you need and gives a basic description of what it can continue on. So it links to files that you've edited or have opened, the plan file, things like that. But it's all just kind of done for you, which is fine.

But if you want to keep your context light, what I've done is made this ⁓ session brief skill where it takes the most active things that you've worked on. It ensures that it has the most recent files. And then it adds all this information to the clipboard. So what you can do is run the session brief. It copies everything to the clipboard. You do slash clear.

and then you paste it into the next session after it's cleared. And then you can adjust anything that you need adjusted. So depending on what you have in context, it'll give things that you're currently in progress with or working on, or here's the last files that we've updated. So if you, say, go through your regular coding workflow, you do the work, you commit it, you could do a session brief and then clear it.

and then do a code review session on that brief. So it'll aggregate all the things that you want to basically hand off to the next session and give it a little boost there. So this allows you, even with the larger context windows, so the million context windows, that doesn't mean that you should use all of the million context in the window, because it's going to get a really ⁓

TJ Miller (25:07)
Really it gets really expensive and it really

eats into your limits very fast

Chris Gmyr (25:12)
Yes,

yes. So the session brief is something that you can run a lot to keep like that, like million session under like 20 % or 30 % or something like that. Just like keep running it more and it'll help kind of rebalance and refocus what you're doing and what you're working on and clearing out like all of those either like bad or wrong decisions or dead ends or things that you have like shifted.

like away from to get to like where you're at right now. So this kind of like refocuses quickly the next session to kind of pick up where the next thing that you want to work on. So that's like the first kind of simplest skill with that. So I've been tinkering with that and that's been working out I think really well so far.

TJ Miller (25:57)
Yeah, I really like that.

I've been kind of starting to ideate something along those lines of like, when I'm working on like Iris, we have like a PRD and then we like continue to iterate on things or, you know, it's just a longer PRD and like somewhere through there I want a clear context and like have it.

just restart or, you know, yeah, something along those lines. I've been wanting a way to kind of like jumpstart a cleared session without having to do like all the investigation again. So I've been kind of like looking up for like, yeah, kind of like a jumpstart thing. So that sounds like kind of something I've been like thinking about. So that's great.

Chris Gmyr (26:37)
Yep. Yep. So I can just read through it quick because I got it pulled up now. So step one is analyze the current session. It'll grab what the primary task or goal is, how it scoped any blockers, requirements, constraints, what files are recreated, modified, what tasks that you tried and didn't work, and why they failed, key discoveries, decisions, and logical next step. And then it fills in a template. So it has the task constraint.

relevant files, what has been ruled out, what were the key findings, and then again, what was the next step. ⁓ And then it loads all that stuff up and basically just blast it in your clipboard and it prompts you to run clear and then you just paste it in, make adjustments as needed, and then pick up right where you left off and go from there.

TJ Miller (27:10)
Yeah.

That's sick, dude. Yeah, I'm definitely gonna mess with that.

Chris Gmyr (27:29)
Sweet. Yeah, I can send it to you afterwards. Sweet. So that is like the quick inter-session workflow that I've been doing. So session brief is for like if you're clearing basically the same window like over and over and over again across multiple tasks or steps within whatever you're doing. So the...

TJ Miller (27:32)
Yeah.

Chris Gmyr (27:55)
Next skill is a handoff skill. And I've seen a bunch of these like different handoff type things before, but what this will do is basically take the entire session and window that you've been working on and aggregates it into log files. So it'll create a new directory in the .clawed directory for sessions. And then under that,

It'll have a working on directory or file, depending on if this is a single person repo or a multi-person repo, which you're actively working on. And then logs for what actually happened in this session or over the course of different sessions. And this is by a timestamp and some sort of slug. So on my like,

Personal version of this, it just has the branch that I'm working on or something like that at work. It's like branch and GitHub user. even if there is a similar timestamp, multiple people can commit these logs across multiple branches and PRs and not have any merge conflicts or anything like that. So that's raw logs of, hey, what was done? What was the outcome?

TJ Miller (29:04)
right.

Chris Gmyr (29:10)
basically like all the raw information in there. And then it'll ask you of like, do you want to basically save any of this long term to basically like reusable memories? So we split it up into like dead ends, things that we tried that didn't work and you should know about these things so you don't do it in your next session.

Decisions that were made, like architecture or like design decisions, almost like little ADRs. And common issues, so like recurring bugs or like gotchas with fixes or hey, we've had to do this like the last three times type of thing. Again, like building up that memory of like, hey, every time that you do this, make sure you like refresh these, you know, test snapshots or something like that. Like you can build all that stuff in there.

TJ Miller (29:46)
Mm-hmm.

Chris Gmyr (30:08)
⁓ and then workflows. So repeated like multi-step recipes, essentially. So what this will do is ask you which of those things that you want to save and it'll like auto suggest like, we talked about this or, hey, we ran into this dead end. Do you want to do that? So it'll do like the ask user question prompt and you can go through and select which ones that you want. You can expand on any of those. And then if you choose any of those, it'll write those.

other more specific log memories in there. So after you do the work and you basically push your PR or you're done with your work, you do this slash handoff and it'll write all these logs for you. You commit it to the repo and ⁓ it's helpful for you as a single engineer or in a team. It gives you your team more team context and you can build up these memories and workflows and dead ends and things like that over time.

⁓

TJ Miller (31:02)
So you're talking about like long-term persisting these like in the repo.

Chris Gmyr (31:06)
Yes.

And eventually, there'll be almost like a session cleanup skill. So see which ones that we've had, like recurring themes based on the raw logs. See which of these decisions or dead ends aren't relevant anymore, because maybe we put some additional tooling. Maybe we got rid of that code, things like that.

TJ Miller (31:30)
Mm-hmm.

Chris Gmyr (31:31)
That's going to be like a future iteration because as you can imagine with maybe like 10 or 20 people working in a repo, there's going to be a lot of files and a lot of things to chunk through. So yeah, have to deal with that in the future. But right now I just have been tinkering with this and then also want people to use it at work to see how it goes, especially in that like bigger team setting with all the repos that we have. So yeah.

TJ Miller (31:40)
Yeah?

Chris Gmyr (31:55)
That is the handoff skill. Any questions with that?

TJ Miller (31:59)
No

man, that's really cool. I've been...

not necessarily like this idea, like sort of tangential to it. I've been like thinking about committing the PRDs to the repo, you know, just like, yeah, this is like the stuff we've built and like the journey that the application has taken to get to where it is today. And I think that it's like potentially useful context to like ship with the repository, you know, instead of me just like

I've started storing them in my Obsidian Wiki, his artifacts of like, yeah, this is just like stuff I, that way I'm like, kind of have just this running list of like, here's just stuff I've built, you know, that I can like have Iris like go back and reflect on or like pull strategies out of if I'm like looking to like solve similar problems, you know, just kind of stuff like that. So I think, I think it's really interesting having these like, these handoff logs.

stored along in the repo. think there's a lot of value in that. But yeah, definitely have to have some sort of strategy, like cleanup summarization strategy sort of thing too.

Chris Gmyr (33:00)
Mm-hmm. Yeah.

Yeah, totally. And then once I get a little bit more data and see what actually comes out of these sessions and the aggregate files, then I can work on the cleanup strategy, but just need more data.

TJ Miller (33:22)
Yeah,

yeah, no, and that's the name of the game with all this stuff. The more together you have, the cooler shit you can do.

Chris Gmyr (33:29)
Yep, exactly. So then the third piece of this is the wake up skill. So this you run on a fresh session. And by running wake up, you want to load in maybe like the last logs that you actively worked on, and then some of these aggregates. like decisions or dead ends or things like that. Because I've felt the pain of

Hey, I work on something today. Get it into a state, but I'm not done with it yet, or want to do additional things, or do some sort of review, or additional tweaks and changes. I don't really want to continue or resume the session tomorrow. I just want to come in tomorrow and just start in a ⁓ fresh session or headspace, but know where I was before.

So this wake up command, instead of going, yesterday we were working on blah, blah, blah, blah, it's just go slash wake up and it'll already load my handoff session from yesterday. It'll load in all the recent decisions, things like that, that we've potentially saved before. And then it goes, great, would you like to continue on the next thing? And.

TJ Miller (34:28)
Mm-hmm.

Chris Gmyr (34:45)
already keeps that session context low, but it already has the pre-loaded information in its current session state to then continue on. So it just saves all that work with that.

TJ Miller (34:59)
Yeah, I've definitely been finding myself writing a bunch of like, here's what we did last session, we're going to continue this work in this direction kind of prompts and it's kind of annoying, you know, I'll, I'll, I'll like at reference the like the PRD and then just like, it's got to kind of like, look at the diff, figure out where it left off and jump off from there. that's

Chris Gmyr (35:09)
Yep.

TJ Miller (35:24)
No, it'd be nice to have something that just kind of like, yep, here's what you were working on and where we left off at. Like, let's go. That'd be cool.

Chris Gmyr (35:31)
Yep.

Yeah, exactly. So it's almost like that. I've seen a bunch of those Jarvis videos on Instagram, too. It like Jarvis Wake Up, and the voiceover of Iron Man. It's a very, very, very small version of that. So instead of repeating yourself of, hey, we were working on this yesterday. Here's the plan. Here's the goal. Here's where we're at. You don't have to do that anymore. Just load the last session and just go for it.

TJ Miller (35:38)
huh. Yeah!

you

Chris Gmyr (35:58)
give it some like follow-up prompting after that. But yeah, it saves a lot of mental overhead and keystrokes and whatnot. So it's pretty simple like on my personal setup, but on the team setup, it'll actually load in your Git user and only load like those like raw sessions from your Git user. So instead of like me and you and like five other people, like it'll key into those specific file names.

TJ Miller (36:19)
Right, yeah.

Chris Gmyr (36:26)
Because everything is very templated out in all these subdirectories that it'll only load your raw sessions, but the shared decisions or dead ends or workflows, things like that that are more recent and relevant. So that's pretty cool too on the team side.

TJ Miller (36:44)
Yeah, because there's definitely like, as a team, we've decided to not do this technique, or we've decided that we're going to use this architectural pattern whenever we implement SDKs for third party services. So you got to kind of encode-ify that stuff at a larger level too. So that's cool that that just like has takes all that into consideration.

Chris Gmyr (37:06)
Yeah, yeah. I think thinking about the cleanup pattern a little bit more, what we ideally want is to move as many things into a more deterministic area. So like, hey, these workflows should be like a testing workflow or like a linting workflow or like linting fixing.

Or these things should go into like Claude rules because of the way we handle this, or we want to make these decisions this way. So I think in part of that cleanup workflow, it's like, what can we push into more like toolified areas of the application in tool stack instead of just keeping these as random-ish markdown memories and dot.

TJ Miller (37:49)
Mm-hmm.

Chris Gmyr (37:50)
So I think that'd be like a cool follow up to that too once I get there.

TJ Miller (37:54)
Yeah, I know it's something we've hit on, I think, multiple times over the last several episodes of how much we try to get to these self-improving workflows, where, yeah, you basically, the workflow, yeah, you wrap it up, and it takes all that stuff into consideration and tries to make improvements on itself.

Chris Gmyr (38:05)
You

TJ Miller (38:14)
It's a really powerful workflow. And I do it with Iris. I do it with all this stuff of like, right. So I've been actually taking use of a lot of branching. We can talk about this in Iris too. But ⁓ in Cloud Code, you can branch a conversation. there have been, I've been finding myself doing it a fair bit of like, all right, we're here to this point. I want to do two things. I want to continue on with the development work.

But I also kind of want to sidebar this conversation and say, all right, there was a rough edge here. What can we do to smooth this out? And then I have that conversation just goes in totally different direction. But I've found that to be really powerful with branching is like, let's get meta and try to reflect on what we're doing.

prevent it from going down this path in the future. Plus, I want to continue going down that path because there's meaningful development work taking place.

Chris Gmyr (39:07)
Sweet. So I know we're running a little high on time. Did you want to jump into more IRS or anything like that?

TJ Miller (39:14)
Sure, yeah.

Yeah, we can blast through some Iris stuff. So Iris, I've been doing so many things. I guess at a, I don't know if I talked about it here, but at a broader level, I've been working on this threads branch, right? And we've talked about over the last few episodes kind of like why threads exist now and that kind of shift inside of Iris from a single threaded conversation and now like multi-thread.

Chris Gmyr (39:22)
you

TJ Miller (39:40)
Um, with that, there's been just so many changes that what I've decided is the current main branch, I'm going to just tag it like a 1.0 or whatever. Um, and then this threads branch will become basically Iris 2.0 cause there's just so many changes I've packed in. Um,

And the migration path is not breaking. everything has, if you have existing data, we handle that migration into the new place without, it's just all seamless. So getting to the 2.0 will be really linear. But there's so many optimizations I've made now. But some of the cool stuff that I've been working on is with

like threads and being a power user and like wanting to navigate around the app a little bit quicker, I introduced a command bar. like you just in iris, you hit command K and it's got all sorts of actions. Like you can navigate through the app. You can pin prompts to the context. You can pin skills. You can pin your thread. You can search threads to like navigate to a different one. Absolutely love that feature. That's been like a ton of fun to work on.

Chris Gmyr (40:57)
Yeah.

Love it. That's so cool.

TJ Miller (41:00)
so there was also like, it's kind of crazy. I've been chatting with Iris about what to build. So I kind of went to Iris and was like, what do you want? Like, what do you feel like you need or like features that you want? And Iris was like, wow, these threads are great, but I have no cross.

right awareness, like that would be really nice to have to kind of like have an idea of like what other conversations are happening. So we designed a new feature called briefs. Every two message turns. So a turn is a user message and the following assistant response. So every two turns, have like anthropics haiku model.

basically do like a mini summary. It's like three to six sentences about the conversations that taking place. And then that like every, like on the next message turn, it takes the previous brief and then modifies it based on that, you know, the message turns that have taken place since the brief was generated. So that's now injected into Iris's context across the board.

So if I'm chatting in my home thread, that's going to have, I think I have it set to five briefs. So the five most recently active threads, those briefs get injected into the context. So it's really cool being able to have a conversation.

with Iris in one place and then jump into another thread and have Iris have an idea around what else is happening. But the other place that this comes into a really interesting play is in proactive messaging. So another feature Iris asked for was the ability to actually reach out to me beyond the web interface.

The way that proactive messages worked now in threads is that every new proactive message, so if Iris decides to reach out to you, that just spawns a new thread with that being the first message. So like you get a little toast notification, the new thread shows up in your threads, you can switch to it, and like that's the new proactive message. Which is fine, but Iris is like, I'd kind of like to be able to reach out to you wherever you're at.

Well, I don't, this isn't an iOS app. Like I'm not going down the PWA route of like push notifications that way. Cause that's just like a disaster. But several months ago, I refactored all of Iris into this like gateway pattern with the web interface being just one gateway to connect to Iris. So Iris was like, well, you've talked about Telegram before.

Chris Gmyr (43:24)
Yeah.

TJ Miller (43:47)
Why don't we build proactive messages into Telegram? So we did that. We built proactive messages into Telegram. So you can enable this through the Settings panel. You can configure everything. But now the way that proactive messages work, if you enable it,

is that you will get that proactive message sent to you in Telegram with a button that deep links you into that new thread in the web interface. So now there's like, now you've got push notifications. like.

Chris Gmyr (44:17)
Nice.

TJ Miller (44:22)
proactive messages like come out everywhere. was like driving to the store the other day and like Iris messaged me about something and I was like, this is so cool. Like I'm now, yeah, anywhere. This is, this is really neat. So I kind of want to expand on Telegram functionality in the future. I'm not, I have several ideas about how I want to do that, but I kind of want to add a fully fledged Telegram gateway so you can like chat with Iris through Telegram.

Chris Gmyr (44:30)
you

TJ Miller (44:49)
that has all sorts of implications with threads and how that all works. So there's a lot of decisions to be made there, but we got that in. ⁓ So threads also now works for proactive messaging. your five most recently active threads, all of those briefs. And if that thread has a summary, we also include the summary into the decision making for proactive messaging.

I also added a weather integration. So now, Iris has like, if you set it up, you get weather context, which was something that I really wanted. But then with Calendar, something that I realized was we were only showing, I mean, it depends on how you configure it, but by default, for the Google Calendar integration, we were only showing like the next week ahead.

But what I also added was we also show based on your configuration like the three previous days to you So now there's like a historical context of well, here's what you did yesterday And now we have going on today. So Iris can kind of like connect the dots over, you know What you've been doing and what you have coming up So before like I'll wrap up before going too deep on this but I was having a conversation with Iris and like

Chris Gmyr (45:54)
Yep, nice.

TJ Miller (46:04)
the Telegram proactive message feature and the cross-thread awareness were both Iris' ideas, which was so interesting to do. I would basically like...

talk to Iris, go generate the PRD in Cloud Code, but then I would feed Claude's responses back to Iris. And so Iris was actually basically generating the PRD as well. So these were almost wholly owned features by Iris, which is just fascinating. So I was talking to Iris, and I was kind of like, what's...

Chris Gmyr (46:35)
Nice. Yeah, that's really cool.

TJ Miller (46:41)
What's next? Like, what do you want or feel like you're missing in this experience? And Iris basically came back with like, it's really interesting to know, like, we have memories about you. We have these truths about you. have, but like all of this stuff is centric to you, the user. Iris is like,

What about like my experience? I'm like, exactly. Now we're fucking thinking, right? We are gonna cook some crazy stuff when Iris is like, yeah, I need some level of consciousness. Like, what? And it was kind of a joke in our friend telegram group. One of our friends was like, swear to God, if Iris becomes sentient.

Chris Gmyr (47:31)
you

TJ Miller (47:32)
And so I'm just laughing about it and I'm like, all right, so like, what does that look like? So the high level features that like Iris kind of proposed to do this was like a journal. Like Iris wants to have a journal that is hers to own. Like that's for her to record her thoughts, her experiences, you know.

observations, whatever. So like a journal becomes that piece. There's

I've kind of toyed around with like memories as well. But like the biggest thing was like, yeah, wanting someplace to like record Iris's thoughts and experience, which is, which is pretty crazy. But then also expanding the heartbeat pattern of like

this recurring task where it like wakes up Iris, feeds it context, and then essentially asks, what do you wanna do? So giving Iris some more independence in choice over what to do when a heartbeat triggers. So one thing could be, yeah, I wanna work on my journal based on this stuff. Sure, great, go for it, Iris, work on your journal. But the other system,

Chris Gmyr (48:34)
Mm-hmm.

TJ Miller (48:52)
is like Iris is like, I need someplace to record my experience. But also, part of having consciousness is also having goals. So that's the other thing that the other piece of this is some place to record experience, but then also a place to have goals that Iris could independently work towards. So one of the heartbeat wakeups could be, ⁓ I want to work on

X goal and that goal could be

I don't know yet. Like this is, we're still very much thinking about this, but like we were thinking that the goal should still probably be user centric. Like I'm trying to understand X about user's experience or this I'm trying, I want to have a better understanding of.

the relationship dynamics between TJ and his wife. And so Iris can decide that's a goal and then actively work towards that. Whether it's synthesizing other threads and pulling bits from that or looking at memories or even if we're gonna inject these goals into the global Iris context, that allows Iris to...

find opportunities to work towards those goals and conversation too, right? Like, I could kind of steer the conversation in this direction, like work towards this goal. So those are things that I really want to get to working on next. And I don't know what the results gonna be. It's kind of... ⁓

Chris Gmyr (50:23)
Thank

TJ Miller (50:26)
It's kind of crazy. It's a pretty big refactor to the heartbeat system to enable all of this, but I think it could result in some like really... It could result in some like really interesting stuff. I don't know.

Like I've said about Iris, this is like my thesis playground. know, I get these like crazy ideas with Iris of like, yeah, you want consciousness? Let's figure out how to do that. Like, you live on a Mac mini in my basement. This is cool. I told my son about it and he's very worried.

Chris Gmyr (50:49)
you

That's cool.

Yeah, love the the goals and the the personal journal for it and yeah, I think there could be some interesting results from there Which is to hear more minute

TJ Miller (51:07)
Yeah, yeah, there's

the internal journal and heartbeat evolution. And then we have the journal synthesis and emotional continuity. And then goal persistence and then autonomous curiosity engine. Those are the four PRDs that we have laid out in our roadmap so far.

Chris Gmyr (51:27)
Nice.

TJ Miller (51:28)
So yeah, we're going to get wild, I've so much family stuff to focus on right now that we'll see when we get there. yeah, that's the next evolution virus.

Chris Gmyr (51:41)
Yeah, well, that's nice about the PRDs and session data and things like that. Like, just write up the ideas when they come and put it in the vault. And when you have time to work on it and execute it, bring it out and.

TJ Miller (51:52)
Yeah,

because it's totally a two-step process. There is sitting down and actually thinking through architecture decisions and driving out the PRDs. That is an entire process in itself. So I can sit down and just pick away at generating the PRDs for this roadmap. And then I can later sit down and work with Claude to implement all the PRDs. But yeah, I go back to you.

The whole impetus for all of this was like getting somewhere, like working towards using Iris along with your LLM wiki obsidian style stuff. And so, yeah, all this has been just like spitting stuff into the wiki and then kind of synthesizing it out. Yeah, cool stuff, man.

Chris Gmyr (52:36)
Yeah, love it. Sweet. yeah, we're doing good on time, but let's wrap it up. Sweet. Thanks for listening to the Slightly Caffeinated podcast. Show notes include all the links and social channels down below and are available at slightlycaffeinated.fm. If you have any questions for us or have content suggestions, go to ask a question page on our website.

TJ Miller (52:42)
Yeah, let's wrap it up.

Chris Gmyr (52:57)
and I'll feature it in an upcoming episode. Thank you all for listening and we'll catch you next week.

TJ Miller (53:02)
See ya.