Reflection Skill, PRDs, and Self-Improving Workflows

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

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

Ooh, I don't know. This week has felt like such a blur. I guess some cool stuff that I did. So my wife earlier this year took my son to his first real concert. And they went and saw Bill Murray. my wife's disabled. So they had ADA seeding. they had like.

pretty solid seats and he was able to like kind of move around a little bit and like he got super into it and was like dancing around the concert. so, I don't know, maybe a, like maybe two months ago, I took my title account and made it a family account and added him because he was like listening to music on YouTube and these ads and just awful. So I'm like, dude, let me just get on title. You can make playlists and you know, so.

Chris Gmyr (00:48)
Mm-hmm.

TJ Miller (00:56)
for the listener title is like Spotify. But in my opinion, it's way better. Definitely higher quality audio, if you're into that kind of thing. But he's been, what I was really trying to encourage him to do is a lot of like music self discovery. Because it's a big deal for my wife and me, like music plays a very important part in our lives.

And so I want him to be able to start expressing himself and finding new music and listening to what he wants to. and so he discovered this band called good kit. and he figured out that they had a concert coming up soon. well, they, they had a tour announcement and he saw somewhere and he saw that like Detroit was on the list. And so he asked, you know, for to go and, we got tickets and.

It was my turn to take him to a concert. So I was all like super stoked to see him get real into it. gosh, he was so low energy. was like, but at one point I look over and I'm like, dude, are you even having fun? He's like, yeah, I'm having a blast, but like super low energy. He was just exhausted because it was also a late concert. Like we normally put him to bed at eight or we're like, we tell.

I don't care when he goes to sleep, that's his own life to manage, but like he's sequestered to his room at eight. Like 8 p.m. you're in your room, do whatever the fuck you want. I'd prefer if he went to bed, but like whatever. it was, doors weren't until seven. So it was just, it was a late day for him. But overall he had a good time. He got some like super cool merch. They had like a tech deck they did.

super cool t-shirts and I had never heard the band before. It was pretty cool. It was a good time. Yeah, that was super fun. And then other than that,

I guess like last week I talked about being super depressed, right? Well, I started some new meds and decidedly decided to like do a few different things to like kind of pull myself out of the the nosedive. I'm feeling a ton better this week, which is great. Yeah, like I'm.

Chris Gmyr (02:54)
That is good.

TJ Miller (02:57)
significant improvements. you know, feels good to be on the lighter side of things for a little bit. than that, just dealing with homeownership. We were just talking before the show, like my basement flooded again. New set of problems. Either it was either a quick fix and we'll see.

or it's going to be a huge construction job. so I don't know. We'll see. ⁓ other than that, like feeling kind of coming out of the depression, feeling creative again, getting some like creative juices flowing. So we talked a little bit last week about Iris and some updates that I was making there. I've continued to make those updates. I've added some pretty cool features to go along with that so we can talk about that later. yeah, man, how about you? What's new in your world?

Chris Gmyr (03:41)
Yeah, and just continuing the AI rollout at work. So some cool things coming out of that, which we'll talk about today. Yeah, and just feel like winding down on the school year, trying to prep and get ready for summer, have some scout activities and some travel coming up in the next few weeks. So trying to get everything wrapped up and ready to go for those. Also trying to think about house projects.

across the board. we had plumbers come out and give an estimate for a tankless water heater, which ended up being a huge fiasco, because we couldn't put it where we wanted it to and had to replace and put in a bunch of new lines. And it was kind of like an optional upgrade right now that we've been just thinking about. So that's going to be put on hold, because it was crazy expensive to do that. So was like, we're just going to stay with what we have right now until it gets much older.

TJ Miller (04:29)
Yeah, oof.

Chris Gmyr (04:35)
⁓ cause we still have like at least five or six years on the tank anyways, as long as you know, nothing happens to it. so yeah. And then, ⁓ doing a bunch of stuff like outside. And I was thinking about getting someone over to run some like ethernet cables, ⁓ instead of doing like full like wifi, like mesh network. ⁓ cause our wifi,

TJ Miller (04:35)
Mm-hmm.

Chris Gmyr (04:58)
at the other end of the house, like isn't super great and it loses connection a bunch, which is kind of weird. But the Wi-Fi system that we have has the ability to not only do like a wireless backhaul for it, you can do a wired backhaul. So what I wanna do is just like have someone come out and throw in a wire that can backhaul from one end of the house to the other and then we'll have like.

two opposite sides of the house that have Wi-Fi available to it and just like a much stronger, you know, wired signal between the two of them. So gotta look into that too, cause with like running Plex and the NAS and trying to do like 4K movies, it kinda doesn't do very well at the other end of the house and like, you know, devices and that's where like the main like TV is and Apple TV and stuff like that. So sometimes there's some like streaming issues when they're...

really is no reason for it besides just it's a long ways away. So I think doing the hardwired backhaul will definitely help with that 100 % instead of expanding the network or getting like a upgraded router and satellite, which is pretty expensive right now as well. So.

TJ Miller (06:05)
Yeah,

that was one of the first things I did when I moved in the house was I ran cat sex to like my living room where the TV was going to be to the room where my office was before I moved it to the basement and then to my room. Just to like make sure that everything had the like, you know, the high bandwidth things like Roku or Apple TV, all that stuff like is not sucking up the Wi-Fi network.

and has its own dedicated things that I can even put like rules around like, hey, you can only consume this much traffic. But I did buy my WiFi router specifically because you can buy mesh nodes. I just kind of, I've got like two of those set up and our house isn't even that big. It's just it's older and like our signals don't go through things as well.

So I've got, yeah, I have to get like two extra nodes on it and it's not so bad, but yeah, I would definitely run a cat six or, you know, across the house and just set up, you know, yeah, I would just do that for sure. Especially if you're trying to do like all the, you know, Plex stuff, like that'll just make the experience so much nicer.

Chris Gmyr (07:13)
Yep, yep, totally. So I had to get someone out for that. But so yeah, just a lot of random house projects and kids things to do. yeah, so nothing too crazy, just life stuff.

TJ Miller (07:27)
Yeah. Yeah. Now that you said it too, it's spring and I got a whole bunch of landscaping projects I need to take care of. yeah. Joy. At least, at least my son was stoked. We walked outside. I'm like, dude, this lawn is getting long. It's got to get mowed. My son was like, yes, I can make money again. Cause I pay him to mow the lawn. ⁓ I told him I'd pay him the same amount.

Chris Gmyr (07:35)
Yeah, nice. Yeah, I mean.

cool. Nice.

TJ Miller (07:53)
that I would pay a company to come out and do it if he wants to do it. So that's his main way of making money. So he's been in soup season all winter. And now he's stoked he can make some money. So I'm hopefully not going to be as alone in yard work this year.

Chris Gmyr (08:07)
Nice.

Yeah, that'll be nice. There's always a lot to do.

TJ Miller (08:14)
Yeah, Sick man. Well, I know we normally talk coffee, but I don't have any updates. I am still just drinking the same old stuff. Although my wife did call me out for not opening the bag of Costco beans yet. So I'm have to get through those next.

Chris Gmyr (08:30)
Yep. Yeah, about the same for me. Just doing the same period of coffee in the morning and then doing the cold brew in the afternoon. So still on that train.

TJ Miller (08:39)
Yeah.

Yeah, I've kind of dropped my cold brew. And so I normally do like a mug of hot, like a thermos of hot coffee, and then I'll do a thermos of like cold brew after. I just like haven't been feeling the cold brew, but my afternoon coffee is now just turned into monsters. So.

Chris Gmyr (09:00)
Yep. Nice.

TJ Miller (09:01)
I don't

know. Yeah. Back on the monster train. But with the way gas prices are, the gas station I buy my monsters from, you get extra points when buying monsters, specifically for whatever promo they have going on. So I've been drinking so much monster. I got $1.45 off of my gas.

Chris Gmyr (09:16)
Interesting.

TJ Miller (09:21)
Yeah, I rolled in on Ian. I was like, yes, this is great.

Chris Gmyr (09:26)
It's a good time to do it. Yeah, I've signed up for like a handful of the different gas station of their like text deal, you know, plans, whatever you give them your phone number and they like text you here and there. But like once you sign up for those, then you can get some money off or if you're like you said, buying something or like ones connected to our grocery store. So I get like fuel points for that.

TJ Miller (09:37)
Yeah.

Yep.

Chris Gmyr (09:49)
So I've been trying to use as many of the points and deals and signing up for those type of benefits and stuff like that too. Because with traveling, that eats a bunch of gas. The new truck has an extended gas tank, so it's 30 gallons instead of 22 or 25. So I try not to get down to less than half because it's just painful.

TJ Miller (10:08)
jeez. yeah.

Yeah.

Chris Gmyr (10:17)
over $100 to fill it up, even just with the mid-grade. So I just try and keep it at a reasonable amount. But it's still a lot of gas, even for half a tank worth. So yeah, sign up for the text deals. It's kind of annoying because you get some of the spammy, like, ⁓ hey, come to whatever to get a free drink or whatever. I'm like, I don't care about your drinks. I just want to get full sense about my gas.

TJ Miller (10:34)
Mm-hmm.

I just want points for gas. Yeah.

Yeah. It's great. Cause like my wife, my wife goes to the gas station that's connected to our grocery store chain. And so we get all the grocery points, like all go towards gas. And so she casts, cashes in there. buy a ton of monster from my local gas station that I go to. And so I get, just get all my points and like cash them in there. like, but I'm

Chris Gmyr (10:58)
the

TJ Miller (11:05)
I'd say like every other fill we're getting at least a dollar off, like with all these points, like definitely worthwhile. Didn't expect to go there, like pro tips, it works. ⁓ So kind of moving forward, I'm really interested about this topic. You pulled up something for reflect and voice skills. So lay it on me, man.

Chris Gmyr (11:14)
Nope, that's all right. Pro tips.

Yeah, so a couple skills that I've been messing around with at work. One is just slash reflect, which is a skill. And what this does is you can manually trigger it if things go a little sideways or you're feeling any sort of friction with what you're working on. It's like, ⁓ this could have been better, or you missed this three times or something like that. So now what you can do.

and I'm going to be putting this in the skills plugins repo, the get work, is you can go like slash reflect and it'll say, and it'll kind of like go through the context. It'll see where things kind of went sideways or you like reprompted it a couple of times. And then it'll try and figure out like what happened and then at what level of the cloud stack can help improve that. So is it a rule or?

⁓ updating the skill that you maybe just called or is it something in the main like repo like Claude file or is it in the global file that this would be better suited in? it'll automatically go through and find out like what it's doing, what went wrong, of like grading it and figuring out like which part of the stack that the edits should be in and then it's just a way for you to reflect on the

session and how it went and trying to build in this like human in the loop review cycle that engineers can go through. Because what I'm seeing at work a lot of times is like people don't know what the capabilities are. Like they know that they maybe can ask some questions or they do do a little coding. But if they're not happy with the results or they're minimally prompting or whatever, they don't always know that they can iterate.

with the AI agent and kind of do that workflow. So hopefully this will be like a little kind of trigger mechanism of like, if things went sideways or you felt a bit of friction one way or the other, just call the Reflect Skill and it'll help you kind of fix it for next time.

TJ Miller (13:32)
Yeah, is this a skill you built or is this one you pulled in?

Chris Gmyr (13:36)
Yeah, it's one that I'm building right now.

TJ Miller (13:38)
Sick. Yeah, because there's actually, I went and just looked for it. There is a Claude reflect skill out there that actually looks really useful as well. So that's pretty funny. No, I think that's really cool. That's something that I would like to get from you to take a look at too, because I think that's very useful.

Chris Gmyr (13:58)
Yeah, totally. Once I get a little bit more solidified and run through it for a week or so, I'll definitely send it your way.

TJ Miller (14:05)
Yeah, I've got

I'll talk about it a little bit later, but I've got one for you to try.

Chris Gmyr (14:11)
Sweet. Yeah. So that's one that I'm tinkering with right now. I'm trying to help with that iteration ability within the team and just, you know, being able to share kind of more of this like harness that I've been working on. So kind of second to that and a new one for this week is voice skill. And what this does, and I talked a little bit about it in my Obsidian blog posts and the content creation

pipeline and all that. Where I have like a writing voice and we talked about this a little bit before too and what you set up for your blog and stuff like that. But this one is more specific to like shorter writing and comments. So I want to leave a comment on a GitHub PR. I want to leave a comment on a JIRA ticket. I want to leave a like reply in something in Slack or something like that. So what this will do

is it's a skill that you install globally. So this will go in the plugins repo as well. But the config for it is not committed. It's something like just in your global Cloud file. So right now, have that in .cloud slash config slash voice.md. And what this is is like a holding spot for all of your patterns, your writing style.

what you want, what you don't want, and then that's a spot where it can iterate over time as well. So what happens is I say something like, want to leave a comment on TJ's PR based on X, and Z. So it'll pick that up automatically. It'll write the PR comment given the additional context and what I want to say. It'll write to a temp markdown file.

TJ Miller (15:35)
Mm-hmm.

Chris Gmyr (15:55)
And then I'm able to pull that into the editor, edit it, and then save it. And basically say, let go or post it. So what that will do is the skill will reread the temp file that you either edited or kept as is. It'll scan the diff of changes and then see if anything is beneficial to save back to that voice.md file.

TJ Miller (16:20)
Mmm.

Chris Gmyr (16:20)
so

that it'll keep on iterating on every post what you changed and what you want your voice to sound like for these comments. So once it does that, like diffing and add to the voice MD, if there's any changes like that, it'll do that. And then it'll post that to wherever. if you have the, know, Cloud comes built in with the GitHub CLI, it'll automatically post there. We have the JIRA and Confluence MCP at work, so it'll be able to post to

Jira tickets and comments and replies and confluence and stuff like that And then it just gets rid of the markdown file So it's again kind of this like self serving self healing loop with the human in loop adding in corrections to what the AI is doing to then better the results the next time that you go through it so I've been focusing a lot on like

How do we keep this loop going for all the different mechanisms and skills and workflows that we have? Because a lot of things right now are just kind of static or things that I have to change in like a skill or rule or something like that. And doesn't always benefit everyone as like individuals on the team using these tools. So like, how do we separate kind of the shared logic and steps?

versus what do we store locally and privately on the individual machines too.

TJ Miller (17:47)
Yeah, man. I've been really kind of getting into this like iterative self-improving stuff too. That's I, this is something that I actually started to recognize the need for too, is that all the writing stuff that I've built for like my voice that it can like draft blog posts for me and come in.

I don't know, relatively close. I, I, what I should have been doing is saving those diffs between like what it generated and then when I ended up posting it as the final blog post. even if I'm having like Claude go in and make the edits instead of like manually do it myself, it's typically when I do a blog post, it's a mix. Like I'll just ask Claude to change a few things and then I'll go in and like just rewrite a couple of paragraphs or something, just like make it feel the way I want.

those diffs would be so beneficial to hang on to you to like then go back and iterate on that, that voice doc. But the thing I started realizing is that is so geared towards longer pros, like building, like writing blog posts or, you know, it just, just longer format stuff. and that's, I think there's totally different guidelines for short form texts. Like, yeah, PR comment, you know,

replying to a ticket or something like that's, I think there are different rules that you kind of want to follow around those things. then I love the, you integrated the like self-improving workflow to it as well. that's something that I think I want to build. What I've actually started reaching for, for short form stuff is I've actually been feeding it my, documentation writing guidelines, because that's,

just geared towards writing clear technical things. So that's done a pretty good job with short form stuff, but I'm probably due to actually make a shorter form one and probably apply a very similar pattern. ⁓ Yeah.

Chris Gmyr (19:37)
there.

Yeah, so just been tinkering with that a bit and a number of different skills and tools.

TJ Miller (19:41)
Love that.

That's really cool, man. I think that's, those are dope. I like those.

⁓

Chris Gmyr (19:50)
Yeah, well, you had a bunch of other things like agentic coding things of a bunch of stuff that you wanted to talk about. So where do you want to start there?

TJ Miller (19:58)
I have so

much to talk about here. ⁓ my gosh. All right. So I think last week I talked about forking chief and Matthias is chief Ralph loop T U I app. I talked about that last week, right? All right, cool. So I've continued to hack on that. I've added a few features since then, I think.

Chris Gmyr (20:09)
Mm-hmm.

TJ Miller (20:23)
I've really gone to town on like refining this PRD, like product requirement doc generation skill.

Chief's PRD generation originally is really pretty good, but I wanted to crank it up a notch. And I think I talked about having Claude go out and do the research on what makes a high quality PRD, and then coming up with a PRD generation skill, then using the skill generator. And then with that, also high quality user stories, because I want user stories to be part of the PRD.

I've been iterating on this like super hard over the last like, I mean, I guess since the last episode, like I've iterated like real hard on it because I've been finding it really useful. I finding I was also using that to improve Chief's PRD generation process. But I also have a like just generic Claude skill for it. And so I've got this like

really crazy PRD generation process now. And what I've kind of baked into it was like, you know, in the skill that Claude built, it really pushes you to create a PRD based on a problem statement. So if I just go in and be like, I want this feature, right, it comes back and goes, all right, well, how can we frame that as a problem statement?

And like built into the skill is like, really forces an interview process. So I literally would go in like with Iris and threads, like introducing like threads to it. I just like, I want to introduce threads, you know, this concept of like context isolated conversations and like, just give it a very like vague high level description. And it came back and goes, all right, we need to frame this as a problem statement. Like, what are you experience?

What's causing you to want this feature? What pain are you feeling? So it asks some questions to tease that out and get you to a problem statement. And then it goes through and the next thing that I really wanted to do was, all right, based on that conversation, go research the code base. Dig in, figure out what affordances already exist, how this feature fits with the system at large.

Chris Gmyr (22:09)
Mm-hmm.

TJ Miller (22:30)
and then come back and interview me again. Like with now that you've done all that research, ask me more questions and like, let's really dig into this. And so like the prompt basically ends with there should be no open questions after this because what I want is I want this PRD to be complete to then like what Chief will do is parse that PRD, extract the user stories and those user stories becomes the tasks that

Chief Ralph loops through. So.

So I just, it just has this whole interview process to like, kind of like go back and forth and like really kind of like tease out this PRD and some like edge cases. so, I made that all compatible with chief and then, then with chief specifically, what I added into this PRD generation process is for every like user story or task, like every, I'm just gonna chief.

uses user stories as tasks, but I'm just going to refer to them as tasks. each task I added on, like the task is complete, now do a review process, like load all relevant, you know, load all relevant skills. So like our Laravel development skill, the PHP development skill, the inertia skill, the Laravel best practices skill, like

load these skills and then do code review on the work you just did. it basically each task now is like double processed, but it's like it does the task and then immediately reviews itself and then goes into this remediation loop of, all right, we're going to do this until the review passes. And then once the review passes, the task is considered complete. And so my goal here was to just get higher quality output.

And what my workflow has been is let me generate the PRD, then have Claude do code review. I've been working on a code review skill, not quite where I want it to be yet, but it does a pretty good job of looking at these specific areas of review. I would have Chief do all the work.

then I would have it do code review and I would have it like dump that review finding into a markdown file too. So now I have a PRD and I have the work that would, the code review that happened immediately after the work was done. So I take both of those, I feed that back into Chief and go, all right, like Claude, how can we improve Chief to like cut down on review churn? Like what can we add to that process? So that's kind of where we got this like additional review step.

you because I want the code review at the end to be as clean as possible, right? I want as little churn as possible when we go to do the code review. So I feel like that's in a pretty good place. I've got some UI bugs. I got to work out on that. But our good friend, David Hemphill, has been working on, I'm letting the cat out of the bag. I asked him if it was OK. He said it was cool.

He's been working on a app called Gent, G-E-N-T. And so I think it's usedgent.com. I know he dropped me.

Chris Gmyr (25:37)
Yep, use JuntaCon.

TJ Miller (25:38)
Yeah,

use gent.com. And so this is like his take on something like Chief where it's got some different approaches to it, but it's a Ralph looping agentic coding application that he's built. we got us and some friends got some copies of it. And so I really wanted to try it on some additional features in Iris to build on top of threads.

So I used Chief to build threads. I got a pretty good feel of that. And so I wanted to, I really just wanted to try Gent. So I've, I iterated some more in my PRD generation process, like actually made that a skill and then just went through cloud code and had that generate the PRD. For this, for, guess the feature that I wanted Gent to build was thread forking.

⁓ so the ability to like any point in time, any assistant message in your message history, from that's in your side of your chat context window, you can click fork and it will create a new thread from of like with everything. It will like basically clone that thread from that point in time at that message. And then you can like have a conversation because like this morning I was working with, Iris on something and then,

I kind of had like a side question about like a side meta question about like what we were doing. And so I like wanted to like fork that conversation, continue the conversation I was having because that was useful in like heading somewhere. But like I kind of needed to have this like side chat too and didn't want to pollute the context. I like fork threading, like thread forking is I...

superpower user like feature, but I'm loving it so far. Anyhow, I got the PRD for this and the way Gent works is like really pretty cool. You feed it the PRD and it has its own task extraction process and it does like task dependency mapping. So you can optimally like parallelize the work because it

it knows what tasks are dependent on the other tasks. So that's, I think, like a super powerful feature. I love the UI, because it's like this animated Kanban board of you see the tickets going from like backlog or like the to-do over to like in progress. And then he's got something really cool, like arrows and animations that kind of show you the task dependency process. And he's got like some cool

settings in there where you can set like review gates. So it's like, before a task is considered complete, it has to run like ESLint, Laristan, Pint, you know, like, so you can define these gates that have to pass before the ticket's considered complete. Which I think is super awesome. I actually have all of that stuff built into my pre-commit hook. I, anytime

Gent makes a commit, it's going to run all that stuff anyways. So I didn't need to use that, but the features there and it works really well. So I had Gent do the work. And it did a really pretty good job. At the end, it was functionally complete. One of the things that I like about the way Gent does its task.

it create test creation from your PRD. It also calls out test cases. So there's like specifically cards calling out for like certain tests, which is really good. And I mean, it was functional at the end of the day, like I got out and it worked well. The sounds like all right, well, like code review, right? So

I do code review on it and it mostly calls out gaps in the PRD generation process. It was like, well, these things didn't account for this thing in the PRD. Like these two things actually contradict themselves. It did call out like some pretty big edge cases that were missed.

So I took the PRD and I took the review and I like took that back to Claude in my like PRD skill or PRD generation skill. And I'm like, look, this is like full of holes. This review was.

The code that Gent wrote was great. For the most part, it found very few issues with guideline adherence that we have in the skills, but it had a lot of holes in the PRD generation. That, you know, Gent can only create tasks based on what's in the PRD.

And like the PRD had like holes in it. so what ended up being like, it gave database columns. So like when you clone, we have to like clone these database columns. Well, it didn't actually look at the Laravel model and look at all the relationships that also needed to be cloned. So while it seemed functional, there were a lot of things that weren't being cloned over to this new forked thread.

So I went back and like, this is that like self-improving loop of like, all right, great. Now we've got this PRD, we got the review. The review clearly calls out like the PRD had holes everywhere. And so I like fed that back to Claude. like, I've got a new PRD generation skill with like more, more things in it now. But I, it was such a pleasant experience using Gent. Like the UI is like really easy to follow.

⁓ watching the Kanban board do its work is great. yeah, I, it's, it's a really cool app. I'm going to actually run a few, a few more things on it. I I talked so much about being skeptical of Ralph loops and like, I still am. And my main concern out the other end is like, not that it's going to create a functional thing. Like, yeah, I believe it will, but

quality control is my biggest problem. And it's awareness of the system as a whole. It'll go in and do its thing to create that feature, but is it cohesive with the rest of the code base? Are we following the standards and guidelines that have outlined in the skills? Because those are important. So I'm now...

in the mission of like trying to figure out like what can I do with some of these Ralph Loop systems to like get pretty high quality code out the other end. And I've found that that PRD generation process is so crucial that I'm putting a lot of effort into this PRD generation skill that

I mean, it's going to be useful for a cloud code, but like it's useful to even like have, if I'm going to be manually doing something, right? Like this just kind of, it teases out some edge cases. gives you solid user stories. like I, one thing I love about Luma is like the JIRA tickets for tech work are so detailed. It's got like very clear. This is what's in scope. This stuff is out of scope. Like here's the feature description.

Chris Gmyr (32:29)
in the

TJ Miller (32:36)
⁓ I'm like, these are great. so I'm like, I'm wondering how far can I get by like taking that Jiro ticket and like, let's generate a PRD out of it, whether I'm going to have Claude build this or I'm going to build this myself. Like, but that interview process that it's, I'm trying to get it to drive you through is so crucial. I

I know I've bit myself in the past by not doing what I'm doing now in this PRD generation process, right? Like, ⁓ it really, it really forces you to think like some of the asks that it's come back with around like what I've given it, you know, to like generate a PRD on like it's questions and it's like digging. I've had to like stop and be like, dude, I gotta go for a walk. I gotta think about like, I gotta think about this and see, you know, a little bit more, right? Like.

Chris Gmyr (33:01)
Yeah, totally.

TJ Miller (33:22)
It came back with some like really interesting questions and like feedback and like, we'll like push back and challenge ideas a little bit. So, I'm super stoked about like where, where that's all heading. I, I don't think I'm in a place where I'm ready to like Ralph Loop at work.

⁓ or like, evangelize it, but I think this PRD generation and like even just vanilla Claude code, and being able to pass it and reference this PRD, like, man, the, the output quality gets so much better. especially for like big features, right? Like this is like, I'm not generating PRDs for like little changes. Like if I've got like a big sweeping feature, I'm

100 % digging into that PR to you.

Chris Gmyr (34:18)
Yeah, yeah, the PRDs that you've shared with me and the iterations that you've gone through it seem really awesome. So I think once you get a little bit more solid, I'd love to try it out. I think I can probably use it on some personal tasks on the side and projects. And David has been sharing a bunch of updates with about Gent with us and looks awesome. So I'm glad that you started using it. So I'm going to.

Hopefully use your PRD stuff and Gent and see what I can do on the side. Because I think, like you said, I'm a little bit leery on bringing too much of the automated stuff to work right now. But I'd love to trial it on my own and see how far I can get and what the output is. So I think all these tools put together are going to be great.

TJ Miller (35:05)
Mm-hmm.

Yeah, I know we're short on time, but man, I'm having such a riot right now with Iris and LLM Wiki. cause like you talked about this stuff, like I stumbled across Carpathi's LLM Wiki thesis, like, and guidelines. And so I talked about with Iris, added threads because I want to do some of this like,

PKM stuff that like you've been talking about building with Iris and open or not when I was with open claw and obsidian I'm like, I need to get My head in order and like I was talking to Iris about this last night With LM wiki and context just like alright, how can we leverage this to? combat like my

specific flavor of ADHD. I'm like, that's how going into this. like, I need, I'm at this place where I'm like really wanting to get my life together. There's like things that I want to do that I need to work towards, which is like notoriously hard to stay on track with all of that stuff.

Well then, my skateboard just fell off my wall and knocked everything off my desk. for the listener, anyhow, threads was born out of wanting to like build that with a lot of the stuff that like you've worked on. So threads was born out of that. then like thread forking was like, I need to have that. I've also added like prompt, like a pinnable prompts per thread so that like that prompt is

added to the system prompt for the entire thread, for the lifetime of the thread. So like it never drops out of context. So I added the LLM wiki to that. I added your blog post to that, Liz, like, pinned prompts. And so I just went in there and was like, yeah, like these are a few things that I want to like capture globally. But like here's my website, here's slightly caffeinated website, here's like Luma's website, a little bit of context about each of them. And I'm like, yeah, just start like.

make it happen. And so it took the LLM wiki stuff, and like, sort of driving out the formatting. And I'm like, I'm really trying to take the approach that you had to have like, you came in with a structure in mind, and then it evolved over time. As they as you learned how the AI uses it. I'm taking that approach like right off the rip. I'm just like, no, you

You organize it. You go organize it based on this, however you want. And I've just started, like last night, just started like yeeting stuff at it. Like here's YouTube videos, here's URLs, like here's some things that I want to research, you know, and just, and it's just filing it all away. Like I went and looked at it this morning.

And it's so cool. Like it's so well organized. I can like find things. It can find things. I'm excited to have it do a lot more with like meeting notes and calendar stuff and daily summaries and you know, a lot of the other things that you have built too. So I don't know. We can talk about that more next week. There'll be a ton more to talk about with it. Cause I'm hot on it right now. yeah. All right, dude, we can wrap up.

Chris Gmyr (37:59)
it's

TJ Miller (38:20)
I'll shut up now.

Chris Gmyr (38:20)
Sweet.

No, you're good. It's all awesome stuff. So we'll get back to it next week. More obsidian fun. Sweet. So thanks for listening to the Slightly Caffeinated podcast. All show notes and links and social channels are down below and also available at slightlycaffeinated.fm. If you have any questions for us or have a content suggestion, 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 next week.