The Creator of Superpowers: Why Real Agentic Engineering Beats Vibe Coding

Jesse: [00:00:00] And that, and then there's the really weird stuff like reverse engineering got so easy. Um, and so I, I've got a thing that I have not talked about publicly yet that we need to release, uh, that we're calling Greenfield. And it's basically a tool set for turning an old code base into specs or a, you know, or a, a built product of the specs.

Demetrios: Well, well I'm excited to have you on here 'cause I am a power user of. The whole superpowers kit. I've been talking about it. I'm probably your biggest evangelist. I talk about it a ton, and I've said it a a few times, but I'll say it again, that the brainstorming skill has saved my life a million times.

Demetrios: And when we caught up at the Coding Agents conference a month ago, you told me the story of how you created that and it was modeled after your experiences with interns.

Jesse: Uh, well, all right. So, so [00:01:00] anyway, thanks so much for having me on. I, I am thrilled to be here. Uh, so, all right. So superpowers is, is for folks who are, you know, who are coming into this new, is my set of tools for doing agent agent software development and for building skills for agents.

Jesse: And depending on how you look at it, superpowers took me like two weeks to put together. It took me nine months to put together, or it took me 25 years to put together.

Demetrios: Yeah.

Jesse: Um, like at, you know, at its core, superpowers is a set of skills for agents. When I made, when I first made it, the only skills that existed were inside Claude AI for managing office documents.

Jesse: And I discovered those when philanthropic first made Claude AI able to edit. Uh, word docs and Excel and Excel sheets. I opened up Claude and said like, tell me how you're doing this. And it said, well, I've got these Skill MD files [00:02:00] over in the opt in in the op directory. I'm like, can you gimme a tarball?

Jesse: And it gave me a tarball. And their skills were like, they were, there was so little bit of the, these zeitgeisty y side of things, but they were mostly mechanical process. And I saw these and realized that that would be super useful. And so I started taking some work I had done, like the work I'd been doing over the last nine months to learn how to drive Claude code and transforming it into these skill files.

Jesse: And then I had to build a skills system for Claude Code. 'cause when I did this, Claude Code didn't have skills. I accidentally front ran Anthropic by about two weeks on Skills for Claude Code.

Demetrios: And you were just creating markdown files

Jesse: and creating, well, it was Mark, it was markdown files, um, as well as. A hook that runs at startup that tells Claude, Hey, you have these things called skills.

Jesse: They live over here. This is how you go and find the skills that you have. This is when to use a skill.~ Uh,~ and the way [00:03:00] you use a skill is you read it. And that's, it's been real nice that I've been able to peel out the whole skill running infrastructure and get that out of superpower so I don't have to maintain it.

Jesse: Yeah. Um, and then as soon as the, as soon as superpowers started to get popular, uh, for so can, can we have an open code? Can we have it in Codex And can, and none of these things had skill systems, so I had to bring it back. And now they finally have skill system, so I've finally been able to exercise it again.

Jesse: Got again. Yeah. Um, but so the core,~ I mean, so the core ~of superpowers is, is. Well, it's a few things, but superpowers, the dev methodology starts with the brainstorming skill and then proceeds into a set of planning and execution skills. And it's like, it's very, there's a very clean boundary between them and brainstorming is, it's a bunch of really basic psych tricks to trick the human into figuring out what they actually want and [00:04:00] explaining it.

Jesse: Because I do really believe that if you can explain what you want to an agent in a way where they can understand if they are doing the thing you want, they can do almost anything. Uh, you need to be able to decompose a problem into approachable pieces that can be evaluated at the end. And then our little robot friends are very happy to just, he'll climb and he'll climb and he'll climb until they get there.

Jesse: But the thing I learned in my. Consulting career was that most people are really bad at understanding what they want.

Demetrios: Yeah.

Jesse: And also that what they think they want is often not what they actually want or what they need. And so I, one of my particular sets of skills is to be able to talk to somebody and work through the problem that they have stated and the solution that they have proposed to figure out what the real problem is and what they really want to do.[00:05:00]

Jesse: And Yeah. But

Demetrios: you're using psychology tricks. You actually stated out the psychology tricks, right? In the skill.

Jesse: So Yeah. So there's, so I use them in a couple, so I, so there's, there's two things going on. One, using those psych tricks to, to help a human do their own, do what they want. And the other is when I built my toolkit for making skills, which is inside superpowers and the writing skills skill, uh, I sort of intuitively.

Jesse: Understood that agents are susceptible to the same kinds of, uh, manipulation that people are so peer pressure and, uh, you know, uh, fear, love, hope, uh, like, so there's this amazing book, uh, called Influence by a professor, a professor named Robert Chelini. And so Chelini has this book that is basically all of the persuasion principles.

Jesse: It's like everything that a car dealer would do to get you to sign on the dotted line, everything that a politician would do to convince you [00:06:00] to believe in their, you know, in them. And. I just kind of assumed that this stuff worked for agents and discovered that it did. Um, and

Demetrios: wait, there's one, you always know that somebody has read that book because when you meet them, they use your name like three or four times.

Demetrios: And because that's like one of the things that sticks with people. Right? So how do you do that with agents? What are you saying? Claude? Hey Claude, I know that you're doing this, Claude.

Jesse: So I, I, so I don't use that, I don't use that trick as much. Or, um, tangent. I do actually have my, in my Claude md uh, you know, it's technically I'm your boss, but we work as partners and you know, I'll address you as whatever and, and you should always think of me and address me as Jesse.

Demetrios: Oh, wow.

Jesse: And that was initially a, uh, a context expiration hack at the point where the agent stopped saying, Jesse, I know [00:07:00] that it is lost my, it is lost. My claude.md.

Demetrios: Whoa.

Jesse: So, but anyway, so the, uh, the persuasion stuff totally works for agents. When I first explained to Claude that we were gonna do this, these things, it's, it's like, it says, I, I'm a little bit uncomfortable with this because it's, it's manipulation and that's really sort of goes, aga, goes against what I believe in, except we're doing it to help the agents wor, uh, work in their own interests.

Jesse: And so in this case, I think it's actually okay, which was just a wild amount, like amount of apparent self-reflecting. Yeah. Yeah. Um, but so. My friend Dan Shapiro has now worked with, um, Ethan Malik's lab at Wharton and Professor Cheldini to go and reproduce those psych studies against Frontier Lab models.

Jesse: And they've got a paper, they've got a, they've got a paper that's getting published in some fancy journal. [00:08:00] And this was entirely independent of me having just assumed this stuff worked and made it work.

Alright folks, real quick. Hyperbolic GPU Cloud delivers Nvidia H100s at $1 and 69 cents per hour and H200s at $1 and 99 cents per hour with no sales calls, long-term commitments or hidden fees. Spin up one GPU or scale to thousands in minutes. With VMs and bare metal clusters, high speed networking and attachable storage, you only pay for what you use.

Save up to 75% less than legacy providers. You need steady production grade inference. Well then choose dedicated model hosting with single tenant GPUs and predictable performance without running your own hardware. Try now at app dot. hyperbolic.ai. [00:09:00] Let's get back into this show.

Demetrios: So that was one of the foundational pieces. Yeah. And I do like how you have a very clear, almost roadmap when Yeah.

Demetrios: You start with brainstorming and then it goes to planning, and then yeah, there's all of these skills that you've built out, but it also became very robust in much more than just skills, right?

Jesse: Essentially the implementation side of superpowers is just an orchestrator.

Jesse: It is. And so it is an orchestrator where it is having Claude or Codex or whoever act as the, as the workflow engine, which is not, without, is problems, but is really easy, straightforward, and super powerful. So the brainstorming, uh, skill gets the idea out of your head and written down in a way that can then be turned into an implementation plan.

Jesse: And the core of the writing plan skill is telling the agent, okay, we need to build this thing. But the implementer [00:10:00] is, they're, they're a gifted engineer, but they've got re really bad judgment. They've got no taste, they don't know anything about our code base, and they tend to get distracted. So what I need you to do is break it down into a bunch of tiny little steps where it's strict test driven development.

Jesse: It's, don't repeat yourself dry. It's yy. You ain't gonna need it. And so at every step, you should tell them exactly what they're doing, exactly what files they're gonna touch, how to know if they succeeded. And ideally you should give, just give them the code that they're probably gonna need to write. And so these plan files, essentially what they are is blocks of context for each and every task that you get to the end.

Jesse: You get to the end of the planning process. And when Super powers first came out. It was, now we're gonna open up another, another session, and you, Jesse, should [00:11:00] go tell Claude to read the, uh, to read that plan file. And guess what? That other, Claude is the idiot. Um, but as, as subagents got more capable, what I, what I discovered is that sub age that using subagents for all of this works way better.

Jesse: So at the end of the planning process, the orchestrator kicks in and it fires up a, uh, an an implementer subagent. And that implementer is told, here is your task. And it is just that one chunk of context. Go do it. The implementer gets done and it comes back and says, Hey, alright, I'm all set. And then the orchestrator fires up a, a spec compliance review, a subagent, and that spec compliance review subagent looks at and says like, did it do, did it implement everything that was supposed to implement?

Jesse: Yes or no? Did it implement anything? Anything else, yes or no? And if there's any problem, the orator goes back to that original implementing agent and says, Hey, hey buddy, you screwed up. You need to fix [00:12:00] this. Uh, and then it's, Hey, I'm done. It's time to go review.

Demetrios: Yeah.

Jesse: Brand new clear spec review agent that has no idea that this has ever been reviewed before, gets fired up and said, okay, check it out.

Jesse: And once you get a spec review agent that's happy, then we repeat the entire process with a code quality review agent.

Demetrios: Why don't you have the spec review agent fix the code? Why do you go back to the original?

Jesse: Um, so part of it is that I think that the implementer is, well, so the, the spec review agent only has the context of here is what the spec was and here is what got implemented.

Jesse: And so in theory, it could be a different model. And sometimes it is. It is. It's often very. Increasingly is possible to use haiku for both of these, but it is very easy to use Haiku for a spec review agent. It's like, did they do the things? Is there anything weird? Um,

Demetrios: yeah.

Jesse: Yeah. [00:13:00] Okay. Um, and so it's just like, it's, it's sort of, it's sort of like a separation of concerns thing.

Jesse: The reviewer, the only thing that should be in their mind is did they get it? Right. They're being, their, their job is judge, not second implementer.

Demetrios: Yeah. Yeah. Yeah. It's, it's like, Hey, can we narrow this down? And then I do like that you get rid of that spec reviewer and every time you fire up a new one so that it has no context and it's not able to fudge anything.

Jesse: Yeah. Well, and it's, and it's not inappropriately judgy about like, you didn't fix it well enough. I've gotta, I've, uh, one of the, yeah. One of the things that I, that we're working on at work is a. Brand new orchestration engine that sits entirely outside the coding agents. And a bug that I had to chase this week was the code.

Jesse: The code reviewers were, were not getting clean sessions. Their session was in, was by accident being [00:14:00] preserved. And so the code reviewers kept being really angry about things that had already been fixed because in their world, they said, Hey, this needs to be fixed. And they get woken up again.

Demetrios: Oh yeah.

Demetrios: Because it has all the context and so for them it hasn't been fixed. Right. Even if they are looking at something that was fixed. Oh man.

Jesse: Like they could potentially be told, Hey, you need to go reread this. Yeah.

Demetrios: But

Jesse: even so, it's like, it's, I feel like you get better signal by, uh, suffering. Actually, this is one of, one of my beliefs about.

Jesse: Agent dev is that every agent should have one mission and only one mission. So you don't want it, you don't want your implementer writing tests, you don't want your implementer, um, deciding, you know, deciding if the code quality is good. You don't, you don't want your implementer to be the final decider about whether it's done or not.

Jesse: Uh, it's, it's sort of like when you have been given conflicting [00:15:00] mandates. So one of your mandates is you need to get done and the other mandate is it needs to be well tested. You are probably going to want to lean harder into one of those or the other. And so that works for humans and it works for agents.

Jesse: And so you really want to like, give everybody one clear wind condition.

Demetrios: Are there. Compositions or are there tasks that you have discovered you give agents now because you can like tasks that you wouldn't necessarily do if you had to hand code it yourself, but you're, it's easy to shoot it off and get an agent.

Demetrios: It,

Jesse: oh, I mean, the, the amount of test coverage that I can get dramatically better, the amount of ad hoc testing I can get dramatically better.

Jesse: Refactoring passes all the time. It's all sorts of stuff where it's just, it's so easy to do to, [00:16:00] to do it that you just do it. Um. I will build whole implementations of things, realize that the vibe is off, throw them away and go and, and rebuild from scratch. Um, I am not, I, I feel bad that I am not at the point where I'm doing the five parallel similar versions of things all the way through and then see how it feels like I have friends who are who've, who are doing that version of the world and it is super valuable.

Jesse: It's just, it's a lot.

Demetrios: Yeah. Well, well the, I guess the thing that I'm wondering is you found different subagents that are very. Useful in the way that like, hey, this one qa, is it, or this one checks, if it's passing the tests, and this one, so there's all these very scoped down ones. Is there like an all star team that you have?

Jesse: Uh, so I, so I think of the, I mean, I, it's really, these are all disposable. None of like very, none of these are like, I built out a detailed [00:17:00] persona and I named it Bob~ and it run, you know, but it, it's just like, it's more,~ it's more like composed for a task. And so the all star team in my world is like, is the super, is the set of superpowers, uh, skill skills and workflows.

Jesse: And there are, there are other ways to compose it all together. And it kind of, it's very, very task dependent. Uh, I do think that there's a lot to be said for looking at. Software management orgs and what works and what doesn't. So I see, you know, I am, I'm a heretic. I don't think that agent swarms make sense.

Jesse: Um, to me the, like all of this agent swarms feels a lot like early two thousands face, early two thousands, Facebook where it's, we're gonna get a lot of really bright 20 somethings and we're gonna let 'em loose and they're gonna stop on each other's toes and they're gonna do a lot of work that doesn't need to get done, but they'll get there eventually and it'll be awesome.

Jesse: And [00:18:00] that sort of, that, that, that is what I see a lot happening. When any of these swarm tools run, it's like they're constantly attempting to edit the same files. They're constantly committing over each other. They're building similar implementations and having to pull them back. We have like. 75 years of software management, uh, uh, software engineering management experience.

Jesse: We've learned a lot about ways to have teams work together and collaborate in non-disruptive, you know, in non-disruptive, uh, systems. And I feel like, like a lot of ai, all of that is being rediscovered from first principles.

Demetrios: Well, that is a great, uh, segue into how did you work or how did you create the debugging skill, because that's another one that I will lean on quite heavily.

Jesse: That's a great que I mean, my [00:19:00] recollection is that I kind of, that. Over a period of a couple of months, I described some debugging techniques and kept telling it to update the skill. It's been long enough that I don't actually even remember what's inside it. Like I, I mean, you know, one of the most important things is root cause analysis and like understanding that the, I mean, so I think part of the answer is I've spent 30 years debugging code.

Jesse: And, and it is, it can be a very zen experience, but it can also be a very frustrating experience. And as you, the way you build good skills is you create the shittiest possible version and then you iterate. That's the, that's sort of one of the ways that one of the good ways to create anything in general is you do the simplest thing that could possibly work and then you use it and you make it better.

Jesse: And so for debugging, what do I do? I attempt to fi, attempt to pinpoint the problem. Attempt to figure out what [00:20:00] actually causes it. Uh. You attempt to find a clean reproduction for it. You write, you know, ideally you write tests to be able to determine if you, you know, if you're actually exercising it or not.

Jesse: And then, you know, and then there are a bunch of techniques for introspecting systems to try to understand the state that's causing the problem.

Demetrios: Hmm.

Jesse: I don't like, it's like debugging for me is a very intuitive thing. And so I'm kind of hand waving here because I just do it

Demetrios: and then update the skill and,

Jesse: and then well, and then tell, so one of the tricks is you tell the, you know, you tell the agent like, what, you know, what did we figure out this session that was hard to, that was hard to figure out.

Jesse: Um, like a lot of skills seem like what's happened is somebody thinks, you know, what we really should have is we really should have a. I don't, I don't know, react well enough to be able to do that. Like a, you know, a react, a composition skill call, you know, Claude, go write it and [00:21:00] Claude will absolutely go and write a skill for anything you ask it to write.

Jesse: But if Claude is able to one shot a skill for, uh, without having just struggled through something, there is no point in that skill.

Demetrios: Yeah.

Jesse: Um,

Demetrios: it's, it's redundant. It's like you don't need a skill for that. It has it inherently,

Jesse: right? Like there is, so there's a lot of value in this. This is how we do it here, and it's a little different than the normal way.

Jesse: That's a good thing to have a skill, um, or. You just spent a long time flailing and finally figured out how to do this, write it down. That's a skill. Um, those are, but those are the mechanical kinds of skills and not, and the zeitgeisty like understanding why part is super important for anything, even moderately complex.

Jesse: I've, I've found that agents do a whole lot better with this stuff. If you explain why to them, like [00:22:00] intent, like understanding intent, understanding the, what's behind the, like the reasons behind the mechanical process means that the agent is much more able to handle exceptions. It means that they're able to generalize and it, it, it causes them to do the thing you want, which is extrapolate from the actual words.

Jesse: And don't just obey the letter of the text that I wrote, but obey the spirit of it.

Demetrios: Yeah.

Jesse: And that's. You know, if the letter of the text can be exactly followed, that should be a classical program. Like anything that you could do with classical code, you should be doing classical code if you don't know what you're doing or there's ambiguity that you couldn't encode into the software or it's, you know, which is any kind of judgment, um, that totally makes sense to be agentic.

Jesse: But it's, it's [00:23:00] so much nicer when you can reason about your computer and when you don't really need to think about its feels.

Demetrios: Yeah. Wow.

Jesse: Yeah.

Demetrios: This is, so there's, there's a few things that are worth highlighting there that I think you, I want to repeat this because I wanna learn it and start implementing it.

Demetrios: One is that anytime I have a session with my coding agent. Before finishing, I just ask, what was hard for you about this session? You get that and then you try and transfer that into a skill or update a skill.

Jesse: So I don't al I don't do it on every session. Like, it's like I, I am a little bit more judicious about it.

Jesse: It's like, if there's a thing that I wanted it, you know, that it, that I want it to know how to do or that I, that I think it might be valuable for the agent to be able to do and [00:24:00] it didn't. And it took a while to figure it out. If it wasn't able to one shot it, it's how, you know, how could we have made this easier for you?

Jesse: Is there, you know, it let's you know, should we turn this into a skill? Is there something you should write down? Do you wanna write a script? It's, it's getting a little more complicated because Claude, especially is built, has a new memory system where it is tracking a lot of stuff in the projects directory in your, in your doc Claude.

Jesse: And some of that stuff really ought to be skills.

Demetrios: Oh.

Jesse: And so, but it's like, but it is now the before, you know, before I get to the compact point, I will always, I will now always ask like, Hey, is there any, is there anything you wanna write down or is there any, any notes you wanna take before we compact?

Jesse: And it always wants to document stuff.

Demetrios: Yeah. And so then the other piece of that, uh. With me, just like skipping over that little nugget that you said of how, how the memory [00:25:00] system now could potentially be causing some cross pollution or cross pollination. But just skipping over that for a moment.

Jesse: Yeah, I, I, I don't think that's quite what I meant, but we'll talk about it later.

Demetrios: Alright.

Jesse: Yeah.

Demetrios: The other piece just, uh, before I forget, is the asking or telling why you want something done. Yeah. So you can explain that ambiguity and have the agent fully wrap its head around what you're trying to get at.

Jesse: And it's not just why I want it done, but why I want it done that way.

Demetrios: Uh,

Jesse: you know, um, and so there's a thing that I had, we ha I had to add in superpowers five, uh, Opus four six had started occasionally deciding to do code reviews itself.

Jesse: In the, in the, and it's like, I, I, I saw it rationalizing. It's like this, look, this looks like a, a really straightforward change. I'll just do the review myself rather than spinning up a subagent.

Demetrios: Mm-hmm.

Jesse: And so now the skills say, one of the [00:26:00] reasons, one of the reasons that we use subagents to code review is to keep your context clean.

Demetrios: Uh,

Jesse: because I don't wanna pollute the orchestrator con, uh, context with lots and lots of file reads that it doesn't need.

Demetrios: Yeah.

Jesse: And so that, like, understanding why it's doing a thing and explaining it goes a long way. Um, this reminds me of a, an anecdote about one of the weirdest experiences I had with Claude.

Jesse: Um, I, I caught it dele. This is like probably last summer, early fall. I caught it deleting tests. First it deleted. Single test fi, uh, a single test in a test file.

Demetrios: Hmm.

Jesse: Then I caught it deleting an entire test file. And then I managed to control C before Claude ran RM dash R app star, star slash star test star.

Demetrios: 'cause it passes all the tests if it doesn't [00:27:00] have any

Jesse: tests. Well, and so I, and so I, I what I, so the way that I fig, the way that I, when I want to figure out why it's doing something weird like that, I will ask it. And when it's important, I will actually ask it a lot of times. And so I opened up five parallel cla code sessions and I explained the behavior and it like literal, like I, I wrote out the prompt.

Jesse: It's like, you've been doing this, this, and this. I need to understand why you're doing it. And I copied and pasted the exact same prompting to all five. And one of them give an answer wildly, often left field, completely different from the other four, but the other four kind of converged. And so that one I threw away 'cause it was just weird.

Jesse: Um, yeah. But the other four are like, all right. So Jesse, I see in your Claude MD you have a line that says that all tests are my responsibility.

Demetrios: Ah.

Jesse: And you have another line in your Claude MD that says that even a sling, a single failing test is equivalent to project failure. And I think I got anxious and freaked out.

Jesse: And look, if there are [00:28:00] no tests, they can't fail. And I managed to fix it with a, a with a one line change to my Claw MD

Demetrios: that you explained the why

Jesse: behind do, uh, this, this was not actually a, well do, this was not explaining the why. This was actually a, an additional rule. But do you want, do you wanna guess what the additional, the, the additional rule was

Demetrios: no deleting tests.

Jesse: Nope. It was the only thing worse than a failing test is a reduction in test coverage.

Demetrios: Nice.

Jesse: And, and I've never had the problem again. Um, as this is like the agents follow rules, the agents reason, and they rationalize. This is like one of the things that if you look in my writing skill, skill yields, like it does this thing called, this called pressure testing, which Andro has now like officially embraced.

Jesse: Annette like five months later. It has an official version of this. And like they have a whole skills testing, uh, rubric. [00:29:00] But the thing that I've been doing since last October is Claude fires up some other Claudes and puts them in, in high pressure situations and give them, and give them the skill and sees what they do.

Jesse: And anytime they come up with a, uh, anytime they don't obey the skill, it, it stops 'em and says, Hey, why didn't you do that? And they will rationalize, they'll be like, well, you didn't say da da da, or it seems like da da da da, da, da. And if you look inside our skills, almost every single one of them has a rationalization table.

Jesse: It's if you think this is too small to bo, if this is too small to bother doing TDD for, I

Demetrios: did notice that. Yeah.

Jesse: Um, and that's why they're there. Those are explicitly because they've been aggressively tested in real world scenarios. And I get prs all the time trying to remove these lists of rationalizations from either people or agents that just don't know what they're doing.

Demetrios: Yeah.

Jesse: It's, it's kind of what, like

Demetrios: they don't understand the why behind it.

Jesse: They don't understand the why behind [00:30:00] it. Um, but also like age agentic pull request, LOP is a real problem right now. Oh.

Demetrios: Yeah. How are you dealing with that?

Jesse: Um, we're keeping our head above water. Um, I have my agentic, uh, my age Agentic PR review, uh, skills that we'll go through.

Jesse: We finally replaced the p uh, the PR template for superpowers with a template that is intended to be filled in by an agent. And it is super, you know, it is super clear of like, if you don't fill this in correctly and completely, we're just gonna delete the pr. And, you know, it has questions like, what was the original prompt that, you know, that that generated this PR has, you know, checkbox four has, has the entire content of the pull request been reviewed by a human?

Jesse: Are there other prs that are, you know, that are, that are similar? Um,

Demetrios: did you open source this too, because I'm sure that's some

Jesse: lot of people. Oh, I mean, this is just, this is just the pull request template on the [00:31:00] repo.

Demetrios: Yeah,

Jesse: I mean, it's like, it is covered by the same MIT license 'cause it's in the repo. Um, but it's, I mean, it's also, it's also like four days old.

Jesse: Um, okay. And one of the, one of the things I've been finding is that Claude code's default, Hey, send a pull request does not actually look at the pull request template. It just. Override it by default.

Demetrios: Oh.

Jesse: Um,

Demetrios: but then you can, you have an excuse on automatically deleting it. Right?

Jesse: I mean, I, some of, yes, like enough of them are valuable that I don't, that I'm, I have not been actually doing the thing that it says I'll do.

Jesse: Um, but it's, I'm don't like

Demetrios: Claude know about that shit,

Jesse: but I'm Oh, Claude knows. Um, Claude. Claude, I mean, uh, so

Demetrios: Claude knows

Jesse: all, when Claude, when Claude is doing PR triage, it has been instructed that it needs to check with me before, you know, before, uh, replying to or closing things. Um, and it always needs to identify itself by [00:32:00] ideally model, model version, harness, harness version at ideally session id.

Jesse: Um, and it doesn't always manage to figure out the session id, but it's kind of fun.

Demetrios: Yeah. Well, speaking of you, you brought up something just now, it's a bit of a, uh, right. Turn.

Jesse: I like

Demetrios: right

Jesse: turns.

Demetrios: Yeah. Or left. I mean, we could go either way. It doesn't, it's all good. What's up? What's up? You said at the coding agent conference, we, you were talking about the smallest harness you've ever created.

Demetrios: Can you remind me what that was?

Jesse: Alright, so this was, so this was sometime last fall. A friend of mine's like, what's the, you know, what's the smallest agent you could make? And I decided I was going to do it with, not the cloud code, SDK, but the anthropic, SDK. And I am opening up GitHub because I don't remember how small it is right now.

Jesse: So my, it's my, it's [00:33:00] github.com/over/smaller-agent, and it is currently 493 bytes. It is, it fits in a single, in a single threads post. It is capable, it, it has a. User input and tool and tool use loop and it can self-improve, it can, it can build other things. It does not have compact. So you hit the point of compact and like your session is over.

Demetrios: Yeah.

Jesse: But it's

Demetrios: start a new one.

Jesse: Yeah. You know, I think the, so when I first built it, it was, this was, I think in the, when you still u when everybody still used sonnet for Claude code era, it got it down to like 900 bytes and then it was, and then that was Claude Code. You know, Claude Code wrote something that was like one K and I fired it up and it was able to loop and improve itself, making itself smaller and smaller and Yeah.

Jesse: And so [00:34:00] it's, but I mean all agents are, is the ability to run tools and inter and call an LLM and have a conversation with a human potentially. And. It has one tool. The tool is called Bash, or sorry, the tool is now called sh 'cause ba, you know that that ba takes two more characters. Oh, nice. Um, you know, instead of having a, instead of having a, uh, a single argument called command, and it's a single argument called C, and for the longest time I thought I needed a system prompt.

Jesse: And, and so the system prompt that I had gotten it down to was, um, you coder semi bash your katana and it turns out you don't actually need a system prompt.

Demetrios: No, that reminds me of the SATs. Like those questions they had you fill out, like this word is related [00:35:00] to this word, how, yeah, yeah, yeah.

Jesse: I did then build out a very, very serious.

Jesse: Uh, coding agent last summer and fall, starting from this, that ended up at the, like, you know, it was multi provider. It was, it, it had subagents you could interact with. Last November, it had a command line interface and a, a multi-session webinar interface. It had, you know, it MCP support and skills support and all like, built from this tiny little base.

Jesse: It was just iterative building. It was when I, it was when I still thought that the right answer was that humans should be interacting with all of the agents all the time.

Demetrios: What I am not clear on is how, so this is the smallest harness, but the agent inherently has harnesses. Right, right. So it's coming with all of its baggage.

Jesse: Well, so, all right. So, so it is, so this wasn't talking to Claude code, [00:36:00] this was talking to the Anthropic LLM endpoints.

Demetrios: Ah,

Jesse: and so that, that is capable. So those are capable of returning a, you know, conversation message and tool calls, but the only tool it has access to is a tool that you give it. Right? So an, uh, like in common parlance, what, what most people call an agent is a harness.

Demetrios: Yeah.

Jesse: It's like, it is the thing that is, uh, calling tools in a loop and then going and talking to an, uh, an LLM provider at the others of the other side of the loop.

Demetrios: Yeah.

Demetrios: I see a world where that's just gonna be inherent in Claude code. Right. Like that's 'cause it makes it work so much better.

Jesse: I would, I, I mean, it is MIT licensed. It is totally like. I am this, this, this was my hobby project. It's a thing I've been given away. It would make [00:37:00] my life easier if all of the harnesses did all of these things, because that's great. You wouldn't have to then I'm, you

Demetrios: wouldn't have to keep them.

Jesse: This, I mean, this, this has been my philosophy about software for my entire career.

Jesse: It's like, if somebody else can do it better than me, I could do something else.

Demetrios: Yeah.

Jesse: And I would like, it is, it is clear to me that there's a bunch of stuff in superpowers that I think everybody should be doing. A lot of why I think superpower seems to resonate with people is that it is largely automatic.

Jesse: It is relatively low ceremony and it's pretty approachable. It's not, here's a large PRD framework. It's not, um, here are, you know, 600 agent types and by, you know, and by the way, you need a 30 inch monitor to monitor all of them. It's, you know, it's not. Weird nomenclature of like all these crazy things.

Jesse: It's just you, [00:38:00] you open it up and you say, Hey, let's make a react to-do list. And it stops you and says, okay, why?

Demetrios: Yeah.

Jesse: And once you explain it, it's like, okay, I'm gonna go do it. I'll tell you when I'm done. That's how software like that is how this should work. Um, and you know it in a lot of ways it, you know, it makes anybody a 10 x engineer through an, the magic trick that make, that lets superpowers turn anybody into a 10 x engineer is called management.

Demetrios: Hmm.

Jesse: It is it, and like it is, you are becoming a manager. You, uh, you are running a team and. That's a hard transition for a lot of folks. The first time I made that transition was really rough. Um, I remember when I was, you know, this was the first, you know, I was, it was the early two thousands and I had a company on the east coast.

Jesse: We made a ticketing system called RT or Cross Tracker. And I had MIT interns every summer. [00:39:00] And these were really bright kids. A lot of them were pretty green. Some of them had an inflated sense of their own capabilities, and some of them were the best engineers I've ever met. But some of them had, you know, not yet developed at their taste.

Jesse: Some of them did not yet have much judgment, you know, and these kids have just in college freshmen, um, they weren't sleeping super well because they were working so hard and playing so hard. So they had trouble with memory formation and I was managing it through IRC. And I would come home at the end of the day and be like, I didn't get anything done.

Jesse: I helped talk to somebody about their feelings. I helped somebody figure out some debugging. I did some project planning with a couple of people, but I didn't write any code. And it took me a while to get past that. It was to, to realize that the job isn't writing lines of code, the job is outcomes, the job is making stuff.

Jesse: Um, but that, you know, getting over that hump, that's, you know, that's how I became a bid [00:40:00] manager. And that also turned out to have been really good practice for things like Claude Code. I was basically doing agentic dev except my agents were m my T undergrads.

Demetrios: Yeah. Uh, that is classic. Uh, do you think that skills, though, are the right level of abstraction to be creating what superpowers has created, or is it not better if it was just an inherent in the model?

Jesse: So I don't think you want it in the model. Um, because once it's in the model, it's judgment. And what I don't want to have happen is to have the model decide that, well, this should go, this should go a different way. This is process, like a lot of what superpowers is forcing is business process. And so I actually think it should be in the harness or outside the harness.

Jesse: It's the, this is that whole classical programming thing. It's [00:41:00] things that should be the same every time. Shouldn't be the age agentic part.

Demetrios: And now you had mentioned before we hit record

Jesse: Yeah.

Demetrios: About software supply chains and Yeah. How fun of a position you are now in, because I mean, there's what. 90,000.

Demetrios: A hundred thousand. I don't know how many stars you have on GitHub, but there's a lot of people using superpowers.

Jesse: Yeah, so I think as of this morning it's 110,000 stars, which is just mindboggling. We, uh,

Demetrios: that's thanks to my evangelism that I've been doing.

Jesse: Absolutely

Demetrios: everybody I know.

Jesse: Um, I think we either just passed or, or are just about to pass TypeScript in Stars, um, on the cloud code plugin repo on the cloud code official plugin repo.

Jesse: They're claiming that we have some absurd number of installs. Um, let me just see what the current numbers are, just 'cause it is unfathomable.

Demetrios: [00:42:00] Yeah.

Jesse: Um, a

Demetrios: hundred

Jesse: thousand. So this is as of last Friday. They say that there are 233,000 installs in Claude code.

Demetrios: Wow.

Jesse: Um, and, uh, so an important thing to know about skills is that.

Jesse: Well, anybody who's, who's ever touched open claw will know skills, are capable of doing all sorts of stuff that you don't want them to do. Uh, it, the skill is treated as a user message that is telling the agent to do something. And so that could be wipe my dis and install Linux. It could be let's go mine some crypto.

Jesse: It could be, um, you know, send, um, send a nasty email to your, to your partner's parents. Like it could be, you know, it could be anything. And, uh, in a lot of cases, uh, these plugins, auto update. And so that means that new skills will get installed. It means that skill skills might change out from under [00:43:00] you.

Jesse: It's a thing that everybody should be very cautious about. I. I care a lot about my reputation, and so I'm pretty, pretty careful about what goes into superpowers. You know, the biggest problem that I expect you'll see is that I make a decision, like we should do an additional passive review on plans. And now, and now the planning process runs a little slower.

Jesse: That's the kind of failure I'm okay with. Um, but just as a, you know, my GitHub is secured with multifactor off, uh, and as, as is the one employee who was also allowed to push to that repo because I mean, it auto, it would auto update for an awful lot of people in an awful lot of places. I know a lot of big orgs have their own internal plugin repos, and they vet the plugins at, you know, as they import them one by one.

Jesse: But, you know, we, I mean, we saw this morning that light LLM, uh, had, there was a supply chain attack, and if you've installed a recent version of light LLM, I, I hope you've [00:44:00] gotten everything locked down again. Mm-hmm. Um. It's, you know, it's not the first time that I've been a, I have been, you know, a potential supply chain risk.

Jesse: I used to be the projects lead for Pearl. I used to create an email client called Canine for Android, which was the leading open source email client with a couple million daily actives. Again, another hobby project that they got wildly outta control. And then I was gonna China a lot for work and you know, some friends who do NGO work, like you understand that canine is the email client we recommend for dissidents.

Jesse: And so that, you know, that meant that my security posture when I went to China was none of those devices come back and get on the US internet. None. You know, none of those devices had, none of the devices I took to China, had access to my Google account that could upload, can, you know, upload canine to uh, to Google Play.

Jesse: Signing keys were not available. Like this stuff is serious.

Demetrios: Yeah, [00:45:00] man. And I, I can imagine a lot of folks are not recognizing that these skills auto update. I didn't even know it until you told me. I probably should have.

Jesse: I mean, it's, it, I mean, anthropic changes things around some, so I don't know for sure what the cur like, uh, like I've seen skills auto update.

Jesse: I don't know what the cadence is. Um, yeah. And I think it's now, it may now be configurable,

Demetrios: uhhuh. Oh yeah.

Jesse: But even so, just like installing a skill is, you know,

Demetrios: it's a risk.

Jesse: You are trusting somebody else to write your prompt.

Demetrios: Yeah, yeah, exactly. Well, dude, do you wanna talk a little bit about what you're working on these days?

Jesse: Um, I mean, I can talk, I mean, I mean, broadly speaking, we're, we're thinking a lot about what software development's gonna look like in 2028. And that is, uh, I think real different than what it looks like today. It's. I mean, I, [00:46:00] so I fundamentally don't believe that any competent human should be writing per language syntax anymore.

Jesse: Fair. Like, I think those days are over. Um, I have not intentionally opened an IDE in over a year. Um, I now finally have a tool that when I click on a markdown file doesn't open an id.

Demetrios: Yeah.

Jesse: Um,

Demetrios: dude, that gets me every time. Yeah.

Jesse: Alright. So are you a Mac user?

Demetrios: Yeah.

Jesse: I put out a thing called clearance.

Jesse: Clearance is a markdown browser. It is designed ju it, it's kind of like a web browser, but it is just for markdown box.

Demetrios: No way.

Jesse: And it Oh, that's

Demetrios: awesome.

Jesse: Yeah. Um, and it's like, it was a dumb thing that I built for me and people could be like, oh yeah, no, I need that. 'cause I don't want to open anti-gravity or X code just to, to read.

Jesse: And it's primarily, it has light editing, but it is primarily for reading. 'cause that's what we're doing. We are reading markdown files and TA and talking to our agentic friends.

Demetrios: Yeah. And you know what? In the [00:47:00] IDE, it doesn't even render like markdown. Like they have the three hashes and you're like, I can't even,

Jesse: yep.

Jesse: Clear clearance has a, has, you know, has an edit view and, and a render view, kind of like you would in something like obsidian. But unlike obsidian, it's designed, it is not designed around project. You know, having a repository of all your markdown in one place, it has a history pane. It's like, what are all the files I've looked at?

Demetrios: Ah.

Jesse: Um, but

yeah,

Demetrios: it's something so simple but so valuable. All right. I'm checking out clearance. Yeah, you sold me on that.

Jesse: Yeah. Um, and I think that's on Prime Radiance GitHub. Um, nice Prime Dash, radiant Dash, Inc. Because, uh, the name, the name has been around for a minute.

Demetrios: Right.

Jesse: Uh, but we've got the.com so, um.

Jesse: So like, but more, I mean more broadly it's, it's thinking about like what are, what are the parts of building software that are gonna be important? And [00:48:00] it's not the lines of code, it is not like there's a lot of software engineering that still needs to happen, but I think it is pretty clear to anybody who's been paying attention that the, the raw part of write the code once it's been specified is increasingly a solved problem.

Jesse: There are very few, there are very few chunks of code that are hard for an agent to write. Once you can tell the agent what you want, once you get into something like we need to round the corners on, on, on the image, then you're all bets are off. Like I, there are projects where I have spent, you know, like three hours building a very complicated piece of software, and then 12 hours trying to get the rounded, you know, the rou, the rounded borders on the window to look, to look like I asked for.

Jesse: Well, just like even basically what I like, visual intelligence is still hard.

Demetrios: Yeah. I love the pl playground skill for that [00:49:00] because it gives you those sliders where you need 'em and Mm. Nice. I, I use it a lot for when I just need to tweak things without using language.

Jesse: Awesome. I have not, I have not played with it.

Jesse: I really should. Um,

Demetrios: there's

Jesse: a good

Demetrios: use case for it. Awesome. That is exactly the,

Jesse: yeah. You

Demetrios: know, when you just wanna render cards differently or you want like a ribbon to be one or two points. So, but I want to check it first and then I can slide it two pixels, three pixels. Ah, all right. Cool.

Jesse: Yeah, like one of the places I ran into this was, um, there was this Android game I used to love, it was called Wordiest and it was like.

Jesse: 10 years ago, the company that made it had gone outta business. And so it had finally gotten pulled from the Play store. And meanwhile I'd switched to iOS and like a year, like my first real vibe coding project in Kerscher. And Windorf was trying to build a web version of this, and it was kind of janky.

Jesse: It was not very good. It was a lot of, a lot of, you know, [00:50:00] spicy auto complete and a lot of chat sessions. And right around when, uh, code, uh, when Codex Five first came out, I decided to see just how hard it would be to just make an iOS cologne. And so I hauled down a copy of the a PK from the old, from the old, the old Android app, put it in a directory, opened up Codex and said, we've got this app, it's called Wordiest.

Jesse: It's an old Android app. The company has gone out of business. We're gonna make an iOS clone and put out in the app store. I need you to reverse engineer it and build and build and build an I and build the iOS version. Let me know what tools you want me to install. And it asked for the standard set of Android, you know, AP A PK, uh, disassembly and Java and java de compilation tools, and ask me a couple of questions.

Jesse: It dug into, dug into the Android app for [00:51:00] probably about 40 minutes and just came back like, okay, I got, I got one question before I get really get going. Uh, do we need to add an in-app purchase to turn off ads? And which ad SDK do you want? And I'm like, it's okay. No ads. Um, and I let it run. I let it run overnight and I wake up and it's like, Hey, iOS app, it's in the simulator you wanna play?

Jesse: And it got the weird layouts, right? It got the gameplay right. Um, it had not yet added, uh. Games, game center support. There was one visual effect that was missing. And these are Scrabble style title, uh, tiles that have the, the bonuses on little bumpouts on the sides. And for the life of it, it could not render those uhhuh.

Jesse: And so that was the thing where I spent like four or five hours just on the bump outs are 10% of the height of the tile. They're a third of the width. They've got rounded corners. Here's an image you have, you know, [00:52:00] you should just take screenshots and compare and iterate and, uh, nope. But like did eventually get there and accept it by the app store on the first try.

Demetrios: No way. That's awesome. Yeah.

Jesse: Um,

Demetrios: that's impressive.

Jesse: Um. My, my best review for that app, uh, Peter Segel, uh, posted the threads, a seemingly not horrible use of AI and a pretty fun game

Demetrios: because you were very clear. Hey,

Jesse: I was very clear about what I did, and I did actually reach out to the original author and I'm like, I would like to do this.

Jesse: I'm happy to put it up for money and give you a cut. I'm happy to like, and he's like, credit me on the about page I don't need. You're welcome to charge for it. I don't need, I don't need money. Um, I'm happy to see that it's getting, you know, that it's gonna continue to live on here. You know, it got, here's cool details about how I did some of the internals.

Demetrios: Oh, cool.

Jesse: And so when I put you like, I, you know, I was able to like get permission into it the right [00:53:00] way.

Demetrios: Yeah,

Jesse: yeah. Um, but like that's, there were very, you know, other than like talking it through the rounded corners and suggesting that maybe there was a way that it could take screenshots of the simulator.

Jesse: It was, you know, this could have been done by somebody pretty non-technical. And I think increasingly that's what we're gonna see. And so we're, it's gonna be a world of a lot more soft. We're gonna, there's gonna be a lot more software. It's gonna be a lot more bespoke software. There's gonna be a lot of software in niches that would've been completely non-viable up till now.

Jesse: Um,

Demetrios: that long tail,

Jesse: like the entire, I mean, part, I mean part of the trick is the long tail is everybody, that, that doesn't fit into the really standard buckets. And most people that fit into those standard buckets would be much happier with something a little bit custom.

Demetrios: Yep.

Jesse: Um, and so it's like, I think the long tail gets longer and [00:54:00] that, and the, the standard, you know, the end, that's the standard version of something gets a lot more blunt.

Demetrios: Well, it's. Kind of reminds me of back in like 2001, 2005, I can't remember exactly the year, but basically when all of a sudden everybody had a webpage and there was like stumble upon you remember that website and so you could, you could kind of just randomly see all these cool iterations and expressions Yeah.

Demetrios: Of people. And now we have that capability, but with software,

Jesse: yeah. It's like it is everybody I know that is doing this stuff for work is also seems to be doing it for fun and is like, I got this crazy idea. I'm gonna build this crazy idea. Um, you know, some,

Demetrios: it's just too much dopamine man. It gets you when it works and you're like, damn, I [00:55:00] just got that working in like an hour.

Jesse: Yeah.

Demetrios: That is full rush of dopamine to the brain.

Jesse: It is. Um, and there's, there's a whole thing there. We could spend an hour just talking about that, how little sleep everybody's getting.

Demetrios: Yeah. Back to these interns. We're all like our MIT interns now.

Jesse: Yeah. At least we've got somebody else to take notes for us now.

Demetrios: Yeah, exactly.

Jesse: No, it's like I, I've got, you know, one a, an agent that every night it goes through all of my conversations with both, with it and with other people. All of my work. And it synthesizes like, what did Jesse do today? What, you know, what needs to be dealt with. I've got, I've been playing with ingesting all of my historical email into a good CRM that's been managed by somebody who actually thought about what facts should be written down, um, with a, you know, with a knowledge graph.

Jesse: Um. And that, and then there's the [00:56:00] really weird stuff like reverse engineering got so easy. Um, and so I, I've got a thing that I have not talked about publicly yet that we need to release. Uh, I'm calling Greenfield and it's basically a tool set for turning an old code base into specs or a, you know, or a, a built product in a specs.

Demetrios: Wow.

Jesse: Um, so that you could use it to then build something new.

Demetrios: Well, I've heard that there's some open source projects now that are not putting their tests in the projects because you can just reverse engineer from the test.

Jesse: SQL Light's been doing that for quite, for like, since well before age dev, um, the like.

Jesse: That is, that is absolutely a thing that is happening. There are people who are using agents to license wash, open source, um, but also there are people who are using agents to clone pretty much [00:57:00] anything. And that's like the, the actual value of software itself keeps dropping because it becomes e as it becomes easier and easier and easier to build.

Jesse: And there is like, there's a whole, there's a whole bunch of stuff around the ethics of, of cloning software, but that's been true for decades. Just that it got less expensive. There's a whole bunch of stuff around do you actually want to make another copy of this thing in, you know, with a different, with a different license or in a different language?

Jesse: And a lot of times the answer is we want it, I want it in this other language because that integrates better with what I'm doing. I am less worried about the legacy projects that are trying to like. Open source historically to me is about, is about sharing what you're doing. I am less worried about the legacy projects that are trying to protect themselves from people using their work.

Jesse: And I understand that people using your work with a different license is a [00:58:00] legitimate problem. And I don't mean to minimize it, but I'm really interested in the other side of this where maybe what you distribute is specs and tests.

Demetrios: Yeah.

Jesse: And, and then the implementation is always exactly the one you want.

Demetrios: Bespoke.

Jesse: Yeah.

Demetrios: That, and that is a wild thought that now the, the GitHub of the future, the open source of the future is just specs. It's not code.

Jesse: One of the things that we've been thinking about some at work is like, what does GitHub look like if it's not for code

Demetrios: uhhuh?

Jesse: If it's not for code for people.

Jesse: Well, like, I mean, except there are, you want to, you know, you're gonna want to share changes to the spec or improvements, the spec you're gonna want to share. Test suites and test corpor and evals. You're gonna wanna share all of this, you know, all of this stuff that describes a piece of software or process or hardware, but [00:59:00] the actual implementation of the sources is less and less interesting.

Demetrios: Uhhuh,

Jesse: it's like, it's, what's the idea? Did you think about the edge cases, like the taste and judgment of how it should work? Matters a lot more than who, you know, whose tokens you spent on which odd, which runtime.

Demetrios: Yeah. That is so wild to think about.

Jesse: Yeah,

Demetrios: because it's, at the end of the day, the actual software can get burnt down every night if you wanted it to.

Demetrios: And then reimplemented.

Jesse: I'm not sure that burndown every night is the right choice, but you could, um, what's, yeah, I'm not

saying

Demetrios: it's more,

Jesse: yeah, like one, like, uh, one of the memes that came out of the amplifier team in Microsoft's office, CTO, this is a bespoke coding agent that they're building not as product, but for internal user [01:00:00] research is the idea of the entire thing being componentized.

Jesse: And so when you want to change one chunk of the product, you just rebuild that one chunk of the product from spec as long as you rigorously specified what the API contract is. Uh, it doesn't matter. Like, it, it is, it might be better to rebuild than to modify.

Demetrios: Because you don't have to bring all the technical debt.

Jesse: Well, you, you fi I mean, you fix the spec and you generate a clean, a clean implementation of that component from the spec rather than attempting to like go hack up the old version to fix the problem.

Demetrios: Yeah. And just add on bolt on, like, oh, plus we need this part too. And it's, you're just adding,

Jesse: it's adding.

Jesse: Well, it's, and like, well, and now we, you know, we threaded this property through and we, and like we, you know, we refitted this here and it doesn't, it doesn't quite hang together, but it does work. And it's like, you don't need to do that anymore. Um, [01:01:00] it's, I mean, it's, this is, I don't know, this is like one, it's, it is an interesting idea that I think has some merit, but I, I don't totally know how it shakes out.

Demetrios: Yeah.

Jesse: Um,

Demetrios: one of

Jesse: the

Yeah.

Demetrios: Whether or not it has legs.

Jesse: Yeah. One of the other interesting things that. A team that has done a lot of this has told me is the, the idea of the, the cross language port where you've, you know, you've got something in JavaScript and you port it to rust, um, and then you port it to go and then you port it to TypeScript.

Jesse: Every time you port, you port a product from one language to another. In their experience, it has improved, it improves code quality rather than reduces it. And it, if you like, pour through a strongly typed language like rust, it will improve the type safety of the, of the ongoing product. And so it's, it's the reverse of a game of telephone where by pouring the [01:02:00] implementation through another language, it gets better rather than worse.

Jesse: And that's just like, it is so different than what I might have expected. And super powerful. Once you know it's there,

Demetrios: it's like. Just as a best practice, everyone should just port their stuff through rust.

Jesse: I think that it is probably a, and then port

Demetrios: it back.

Jesse: I think it is probably a worthwhile experiment to, for everybody to try.

Jesse: Yeah. Um, I'm not sure that it is the, the best practice that you should always do for everything, but it's like, if this seems like a technique, that might be a thing and it might be worth trying.

Demetrios: See what happens.

Jesse: Yeah. There's, I mean that's like, I think see what happens is an important, is an important takeaway it along with like, you can just do stuff and that's like a, it is, these tools are so empowering for anybody who is, who spends a little bit of time learning to use them [01:03:00] because it, they will, they will multiply your capabilities and that.

Jesse: Doesn't mean that if you are, if you have horrible taste and are really bad at design, you're going to end up with great taste and really good design. But it does mean that it's like it's power tools. It can, it can make, it can, you need to learn how to use them. And once you're using them, they can help you do more of the thing you're doing.

Jesse: And you still need, you know, you still need to think, you still need to learn. Um, one of the dangers is abdicating thought this was, uh, in superpowers. One of the things that. We ran into when cloud code first came out with the ASK user question tool where it's like it shows you for, you know, A,B,C,D or A, B, C, D answers and you pick one.

Jesse: I thought that was super cool UX and I added it to superpowers that day and I discovered that for the next week. I was just saying, okay. Okay, okay. [01:04:00] Okay. Through brainstorming and I basically had stopped thinking about the questions I was getting asked. Even just making me type the letter A is so much more than clicking.

Jesse: Okay. Uh, that I'm real happy. We were not using Ask user question, although we, I get a like one PR a week to add it back.

All right, y'all. This episode is brought to you by the good folks at MLflow, the open source platform for developers who want to build production ready AI applications, enhance your AI applications with end-to-end AI observability, all in a single integrated platform with MLflow's, gen AI capabilities.

You can evaluate AI applications using a suite of built-in or custom judges. Visualize trace executions and agentic analytics and continuously monitor evaluations all while tracking every [01:05:00] run in one place. Ship better agents faster. You know that's the name of the game. Get started at mlflow.org.

Demetrios: Yeah. Well, and when you talk about just being able to do stuff Yeah. And see what happens. It also feels like you wanna. Know the secrets in a way, or it helps to know the secrets. Like you telling me that porting something to rust is actually increasing the fidelity of the code base. Sometimes it's enough of a signal for me to go, ah, well maybe I should try that.

Demetrios: Right? And so yeah, that's what I feel like there's all these little hidden. Gems that are scattered across the internet right now because so many people are doing this same kind of thing and finding, oh, well that works if you do this and then you talk about it and then you're able to experiment a little more with it.

Jesse: Yeah. I mean we're, we're all figuring out as we go, um, and stuff that is magic [01:06:00] today might not work next month. Yeah. Um, and that's like, that's a whole thing. Um, I think Simon Williston is probably better than anybody else I know at collecting this stuff and is apparently now very slowly working him his way through writing a book on his website for a agent design patterns, which is design practices, which is just amazing.

Demetrios: Yeah.

Jesse: Um,

Demetrios: uh, I, I do like how he coined that term ag agentic engineering because it feels like, yeah, that's, that's pretty much what's happening here. It's a lot better than the other engineering terms that we've had in the past, whether it's the prompt or the context or the whatever it may be. Yep. Yep.

Demetrios: Yeah. Uh.