Try Harder. Ultrathink!  ## Metadata - Author: [[The Changelog: Software Development, Open Source]] - Full Title: Try Harder. Ultrathink! - Category: #podcasts - URL: https://share.snipd.com/episode/0e040ed0-ce54-44f9-83e1-78a266b98064 ## Document Notes ## Highlights - **Bug Fixing with AI Assistants** - Nick fixed a bug by having Claude work while he focused on other tasks. - Within an hour, Claude diagnosed, fixed, and prepared cleanup queries for production. Transcript: Adam Stacoviak 30. Nick's in his, his, uh, dinner robe, post dinner robe. You know, he's got his slippers on, pours himself a nice glass of wine. Yeah. Or bourbon. I'm thinking wine is more Nick style, but maybe he'll correct us. Cracks open the old, you know, sits back, maybe lights up a cigar. Cracks open the old laptop and just goes perusing dot files, baby. Nick Nisi You forgot the George Michael in the background. Adam Stacoviak With George Michael. Yeah, there you go. Is that accurate, Nick? Or are you more of a bourbon guy than wine? Nick Nisi I'm more of a sparkling water. Okay, fair. Sparkling. Adam Stacoviak Okay. I was trying to really, you know, play it up, but sparkling water is probably. He's finally at the Topo Chico, Jared. You got any Topo Chico in you? I have some upstairs. Jerod Santo I have some upstairs. And right next to me. Upstairs upstairs hide it in the closet oh this is awesome oh good stuff well i have officially installed claude code just so you're there you're with us i've officially installed and I have not played with it officially yet because i'm just i'm a waiter you know i'm i wait until things settle down a little bit i had to wait till you know really convinced me to do so i've Just been playing with things in the cloud really you know copy paste i've been the api for a bit right i'm gonna let the agents be the api and do the copy and pasting for me jared you've been Adam Stacoviak Using it for a while right yes i've been using it for a while but still just in like out of the box mode. Like I'm not tweaking and customizing, but I've definitely been experimenting with getting more out of it and just having fun. That's what that episode with Steve actually did for me was like, he convinced me that this might be fun. And he gave me actionable steps to go try. And I just started having fun for the first time versus being angry or scared. I was just like, no, this is fun. Like I said yesterday, I realized I had a bug somewhere in our production site. It was like people without memberships could somehow create feeds when they're not supposed to, even though I know I wrote the code to make that not possible. But I'm doing Change Dog News. I got all that to do on a Monday morning. And so I'm like perfect use of like, let Claude go work on that while I work on this. And it was just fun just watching someone else do the work. And then I'm like, no man, go look here instead. And he's like, okay, cool. This works. You know, he's like, can you give me a, a feed ID of something that, of one that actually, you know, shouldn't exist. And like, I'll go find that, give it to it and then just leave and go back to my work. And yeah, in a matter of 45 minutes to an hour, had the bug, had the actual diagnosis, had a fix, had tests in place, and then had given me a list of feeds that shouldn't exist and a SQL query To actually go delete them out of our production database. And I'm like, this is cool because I'm not actually working. ([Time 0:23:01](https://share.snipd.com/snip/bfa286c8-fadb-49f3-a789-d13e308c7b0d)) - **Organize AI Tasks with Git Worktrees** - Manage multiple agents in separate Git worktrees to work on concurrent tasks. - Use Markdown plans phase by phase to keep AI tasks organized and track progress. Transcript: Adam Stacoviak Like, cause you, I mean, you code for hours, right? Okay. Nick Nisi Theoretically. Adam Stacoviak Depends on who's listening. No. Well, because I have other stuff I'm doing, right? So like for me, I'm writing a newsletter and I'm recording a podcast. And so I'm having to do stuff while I'm not working on software. And I wonder what I would be doing if I was like, if my job that day was to fix that bug, would I be on Reddit? Would I be bored? Would I be off doing something else? Like when you're using these things throughout a day, are you also coding in a different area of the code base while it works or you how do you manage because there's a back and forth where Nick Nisi You're not really doing anything oh yeah sitting and waiting is that's like the amateur right like there's a process that you go through and like you know getting to these tools is like You know the a good you're you know you're in the fourth dimension now but you're just like twiddling your thumbs in the fourth dimension. Now you have to go to the fifth dimension where you're like going through, you, you've got multiple of these agents. I think that's something that I really took away from, from that podcast with Steve was like, just, you know, set these agents off to do a bunch of different things all at once. And like, you're no longer involved at that. You're a manager of agents. And maybe you have an agent that manages sub agents and goes from there. And like, yeah. Um, I, I posted a link in the, in the things it's a blog post I have called how I use get work trees. And I only call this out because it's by far the most popular post on my website. Um, even still, and it gets like regular hits. And I think it's because of AI. And I think it's because work trees are just a fantastic way to work with these tools because you can set off Claude in one work tree to do one task and then have another work tree where it's Doing another task. And these are effectively just like different checkouts of the repo, right? But they're all connected. And so then you can just like hop between all of these and do like five things at once in the same repo and then come out with five different pull requests. And that's really easy and really fun to do. And like what I'll often do is like, just because I want to feel like I'm still hands on keyboard a little bit is like, I'll have it off going and doing stuff. And then I will work like more in depth on one of the problems while I'm like kind of traversing between TMUX sessions to keep the prompts going for other ones. But the main thing that I do is like, I make it, I make it write a plan and stick to it. So I'll spend hours at the beginning going through a Markdown plan. And it's just like step-by phase by phase. They like, it likes to do things like phase one. And then there's like steps one through five of phase one. And then that way, like I can focus it on doing that. Like I know exactly what it's going to do from start to finish. And that's written to a Markdown file. So it's like in there, it's something tangible. I can, you know, kill this instance of Claude and come back and just give it that, that file and continue going. And then I tell it like, all right, let's work on implementing phase one. And it'll only do that. And then it will like, you know, do the changes and it will stop and think and ask questions based on that. I have like a whole bunch of prompts that I, you know, tell it to ask me questions or wait on things. And then it will, it'll stop right there. And then once we're done and I like the phases, I usually try and keep them to something that I can like manually test to make sure, yes, this is actually working correctly. Then I'll do a code review and fix things or, you know, clean things up. It likes to be very verbose. It's not very dry. Not that it has to be very dry, but like it does things that I probably wouldn't do. So I try and make it do things that I would do. And then, and then you can do like a slash compact or a slash clear. I usually just do compact and that will like compact the session down and give it a summary of what we've talked about so far so that you have a new context window full of, you know, fresh Memory. And then now we'll start on phase two. And it kind of has like hints of what it did in phase one and it can go back and check, you know, obviously. But then it kind of starts fresh with phase two and goes from there. And that's how I like I really like to do it. So like I set it off to do phase one, I go work on something else and then I come back and see where it's at and, you know, monitor it. But, uh, I'm usually doing like two or three things across two or three repos, usually like just like all over the place. I maintain a lot of SDKs. So I'm just kind of constantly jumping around. ([Time 0:42:50](https://share.snipd.com/snip/62ee93c1-17b3-46f1-b528-823ada39ff2e)) - **AI Adoption Curve in Productivity** - AI can seem to slow developers at first due to experimentation and adapting workflows. - Personalized prompt engineering and use will define future efficiency. Transcript: Adam Stacoviak That being said, Nick, you seem pretty confident that your perceived speed is actual speed. And curious your reaction to that number. 19% slowdown seems like a bad deal compared to what everybody has been telling us. Nick Nisi Yeah, I believe it. And I also think I'm more productive because of AI. Um, it's, it's because, well, I'm a team of one right now. So like I rely on conversations with AI, like top to bottom, it could be having a do code, like write code. It could be just having an architectural level discussion to flesh out my ideas before i go bother a real human you know i've knocked out all the low-hanging fruit i'm going to sound super Smart because i know everything because i've already had this conversation right and now i'm validating whether it was just you know being telling me what i want to hear which they love To do and i hate it so much um but like we're all the the, like, I'll be honest, I didn't read the paper, so I don't know exactly like the parameters and things that went into it. But if it was early 2025, that's going to be different than June 2025. Adam Stacoviak For sure. Nick Nisi But also like, we're all going through this right now, like figuring things out. It's like, um, we're just, you know, throwing spaghetti at the wall and seeing what sticks, what works, what gets us there. ([Time 0:52:22](https://share.snipd.com/snip/0379dd13-96cc-4ede-ae26-a9fcdab1e68f)) - **Context and Taste Trump Thinking** - Human context and taste differentiate us from AI, more than pure thinking. - AI's thinking mimics but lacks contextual judgment and personal taste which guide decisions. Transcript: Adam Stacoviak So I'm not necessarily disagreeing with you. But I don't know, Adam, I feel like maybe you're anthropomorphizing these things too much. Like you're assuming, well, you say like we think they're on our side or something and you're getting very much in the T2 territory here. I don't think there's sentience. I was decompressing a longer version. I mean, I was saving us from one more monkey story. Okay. Okay. Well, thank you. Save us from as many monkeys as need be. So I don't, I don't think that at a certain level of software quality that can be objectively tested as best of our ability. Produced on our behalf by AI and agents into the future. I don't think not having to look at the code or know what the code does, whatever I said, is all that big of an issue. How is it different than assembly code, which I've never looked at in my entire life? ([Time 1:07:40](https://share.snipd.com/snip/547e0620-2f41-40b7-b0dc-740927e1f19e)) - **Automating the Assembly Line of Coding** - Much of coding is assembly-line work, not craftsmanship. - AI helps accept this reality and automates routine coding tasks to accelerate delivery. Transcript: Adam Stacoviak Drove a stick shift today no when i drove to drive no at one point in my life i drove one once well more than once on a few occasions my friend one, and so I had to drive his. And then one of my early jobs had a stick shift. And little arrogant Jared was too egotistical to say, when someone says, can you drive a stick, my ego was so big, I just said yes. Oh, gosh. And I had never done it before. I had watched people do it. And so I had to just figure it out on the road, which was a terrible day. But that's arrogance and youth are eternal companions. Anyways. Nick Nisi You're able to churn out code now um and you're you're you know potentially looking at it less uh you're definitely looking at parts of it less like just trusting yeah it just works right That is like it's making me realize that for the most part like the code i write the code i get paid to write it's pretty boring the assembly line right this is not handcrafted custom furniture This is an assembly line and i got to get it out the door. Adam Stacoviak Right. Nick Nisi And it's helping me to like come to terms with that. It's been there all along, but it's just like, yeah, that's true. And ([Time 1:13:13](https://share.snipd.com/snip/9764b8d6-57a1-4747-8df2-2affc8f69710)) - **Master AI Prompt Incantations** - Learn and use specific prompt incantations like subagents and ultra think for better AI results. - These prompt keywords act as a programming language to command AI logic and reasoning. Transcript: Nick Nisi That get us to where we want to go and like you if you know these incantations you are casting these spells you're an engineer prompt engineer yeah that's crazy man you like that i actually Adam Stacoviak That makes me mad. Jerod Santo I just wanted to try hard every time. Adam Stacoviak Like, I just shouldn't have to tell you. Like, now I'm talking to a human again. It's like, come on. I shouldn't have to motivate you. Jerod Santo You know, you can do this, Claude. Go and do it. Try harder. Nick Nisi It's like a programming language that isn't a programming language. Like, you have to just know these keywords and do them. And you, like, you can read them through the docs probably, but like you learn them through trial and error. Well, trial and error and just like legend that people pass. Someone else telling you something. Adam Stacoviak Yeah. Yeah. That's a, that's a truth that I think goes, is trending downward over time. ([Time 1:20:49](https://share.snipd.com/snip/8511a302-3ff7-4c72-b2b7-093a2682b240)) - **Why AI Slows Some Developers** - AI slows open source developers due to high context knowledge and time to provide it to AI. - Experts may find AI less efficient than doing the work themselves, causing a productivity lag. Transcript: Adam Stacoviak You can offer that to clean up ai yep for sure there's an opportunity there for now until their code is way better than ours yeah and um and we'll see if that actually happens. But we didn't really rebut this 19% slowdown very much. I mean, we're over here talking about how amazing everything is, but like cold hard facts are saying, nope. Now I think there's a few factors that play into that. In fact, we'll link to this blog post called AI slows down open sourceers. Peter Nauer can teach us why because it goes into details on why this particular set of people in this particular place some good reasons why they might have been slowed down whereas In other areas and other contexts you actually are sped up like Nick confesses to me. So, um, we'll just leave that in the show notes for people to read. ([Time 1:25:07](https://share.snipd.com/snip/9bb34d70-0da2-46b7-8d4c-53150c1561fc))