The AI Sandwich, Agent Primitives, and Designing with AI

When agents can write the code, where should we spend our time as builders? Today’s guest calls his approach the AI sandwich. The bottom slice of bread is humans deciding first what to build. In the middle, agents handle the implementation. On the other side, humans polish the results. Welcome to Agents and Engineers. I’m your host, Dan Gerlanc. Today’s guest is Trevin Chow.

Trevin is a product leader and builder with more than 25 years of experience in product and design, Including leadership roles at Axon, Sketchy, Nike, and Microsoft. He’s also a core maintainer of compound engineering, a plugin for Claude Code, Codex, and other agent harnesses.

In this episode, we also discuss how Trevinn, starting with a product’s goals, uses agents to explore designs before generating prototypes, what it takes to build and test skills that work across different models and coding agents, and why sharing documents and context between agents is still difficult, and how Trevan’s open source project Energon addresses it. Enjoy the show.

So you have a bit of a different perspective on AI agents, AI engineering. You have a CS background, but you’ve been a PM largely the last twenty plus years. How has that shaped your opinion on where we are today and

how you’re doing things.

I think

there’s, yeah, it’s I I didn’t expect to have a different opinion on this like than most. I you know, when I was a kid, I like I you said I had a CS degree, I was like coded when I was like a little kid. but now is like the best time I feel like to work in software. It’s like the craziest time. There’s so many tools, there’s so much stuff you can build. It’s like you’re you think it, you can go build it. But I think there’s two perspectives that have emerged. It’s like

That perspective or one of like a doomer perspective where it’s like making software sucks now because all you do is press enter and Claude Code or Codex or whatever your thing of choice is. And I don’t prescribe to that. Like I’m like of the optimism case in this. And I’m like never been more excited for tech in my life. I I think. this period is just so exciting. And I think the difference is I think maybe it’s where I came from, where I spent the last 20 plus years, like leading product and design teams, like.

I’m a product person through and through, and I r I want to ship like awesome products that solve problems. And it’s never, I don’t identify as an engineer, like a traditional engineer. so I think that this rise of AI and a kind of agentic coding has been one of wow, we can solve so many problems so many different ways, and sometimes much faster. that gets me excited. But I think that’s at odds with the the

Crowd that is like, you know, hates the pressing enter piece of it. I think there’s way more than that. I feel that’s overstating the pressing enter part. But but yeah, so I fall on the completely the opposite side of this. And I think that’s less about the making of the software versus what is it that we’re trying to do with it. And that’s where I have always seen kind of and been excited about technology.

I think

it’s gonna bifurcate a bunch, right? I mean, I think there’s some people you either, you know, you either stay in the game or you tap out. I feel like there’s a mushy middle in which people will stay in because they have to, right? We work for money to sustain a lifestyle, our family and things like that. But at the end of the day, you either have to, you know, show up or kind of get out. And so I think that over time the their attrition will happen and people will either come to terms with the fact that

The nature of the job is just really, really different than what it used to be. And I think the irony is, when you think about it, is the tech industry has spent so long snubbing its nose at other industries, whether it’s trade’s craft or blue-collar work. And it was fine when tech was disrupting other people’s businesses. But now when it’s disrupting our own, I think there is a bit of NIMBYism that’s happening where it’s like, hey, no, this is not.

This is not what I signed up for. And I think that that transformation is happening, whether or not we we want it or not. I think we’re well past the time at which is AI going to be a thing. Like it’s a thing, it’s here, it’s not going away. And I think the real important thing is just like how much of your job is changing because of it. And I think that engineering is getting hit really, really hard for a lot of good reasons. And I think we’re like most transformational things.

I think things are gonna come back a little bit more to center. You know, we’re not gonna be so obsessed with sh one shotting 3D games because 3JS can now do amazing things with Astra. so I think I think that is what’s gonna happen. We’re gonna have this kind of feeling of this tidal wave of change that’s gonna happen, a lot of strong feelings. But I think people are gonna have to learn to adapt if they wanna stay in this industry, w wherever you are in the industry.

totally. I mean anyone who’s

who anyone who’s worked on real software, even if they see like, you know, the call of duty clone that was done with a you know one shot prompt with, you know, whatever what whatever model it was, you cannot do that with an existing product and just say like, redesign the product, make it good and make no mistakes and profit. Like that just it’s just not it doesn’t work this way. And so there’s more craft in it and it’s a different craft.

But it’s there’s more craft in that in how you apply the, you know, agent intelligence to even do design. And I think the squishier the topic is that requires more subjectivity, like design and user experience. It’s it’s gonna take, I think, more turns of the crank before kind of AI really sets in to there beyond the flashy, you know, demos that we see on Twitter. and why we see so much.

disruption of, you know, writing code about, you know, you know, case sensitivity and checking bounds and security vulnerabilities and things and bug fixing. Like all that feels so like, you know, digging trenches now if you kind of are embracing AI. And I think most people would say, if they’re being honest, there’s a lot of work that engineers don’t love to do and they’re pretty excited they don’t have to do it.

But they’re f afraid about the stuff that they like to do and they’re afraid that of that being taken away. So I think the clutching of the pearls of that piece I think is where we’re currently in this this phase right now.

think there’s a place for even like the hardcore side of things, like there’s a software engineer, Daniel Lemire, I may not be pronouncing his last name correctly, but he’s written a lot or worked on a lot of very low-level libraries.

Like simdjson and other things that are used in like browsers and databases. And he had posted recently about how, in certain cases, AI changes to these have sped up certain parts of these libraries, like 6X or things like that. And these

Yeah, there’s just certain work that I think is optimized for a certain kind of, you know, agent right now, or that or maybe vice versa, the agents are optimized for the certain kind of work. and that’s just gonna continue to get better and better and better. And so it’s like I think the main the biggest thing I think about is it’s a transformation wave as cliche as that sounds, and just it depending on where you are, it’s gonna hit you in some way. And

the strength of that you’re gonna feel differently when you feel it’s gonna be different, but it’s gonna come for you. And so I think the biggest thing that I always think about is like the people who I see most successful that I’ve worked with in this transformation are those that embrace the change and are like down to try, down to be challenged, down to, you know, not feeling like the expert still about everything and kind of have a learner’s mindset. And I think that’s the the key way to be successful right now.

It’s interesting. I think about it the same way as with writing. It’s a creative endeavor and it’s a point of view. At the end of the day, you know, a design, whether it’s visual or user experience or or you’re writing fiction or nonfiction. Like there are creative endeavors in some form or fashion that are taking a point of view from a person or which is shaped from the goal that you’re trying to accomplish. And the way that I see it is it helps you sharpen it. And depending on where where you are, and I, you know.

There’s been a long time where designers, you’ll s you have seen specialization where someone’s a vi someone used to just be a visual designer and they’re used to be a user experience designer. And then you would pair them up and they would kind of like team on projects. The most common thing I hear from designers now is that it’s they’re able to be much more multifaceted because AI tooling has enabled them to make up for some of the hard skills that they don’t have or to make some hard skill that they have much better to compensate. And so

I think I see it as a way to amplify someone’s point of view. And I think a lot of it you hear about human judgment and human taste. Like ultimately, that is what I’m saying without trying to sound so cliche about it. Like people bring a point of view and they they they they know what they want, but they don’t know exactly how it should look. Just like when they’re writing a, you know, article, an essay or whatever, they they have a point of view, but they don’t necessarily know how they want to express it in the right way. And I think there’s

You know, when you think of the written word, there’s multiple an infinite number of ways to say to get your point across. And some ways are better than others. And so can AI help you in that way? And I think the same way for design. And I I I think the biggest one of the biggest parts of the disruption is the speed in that blog post that you kind of are are referencing. My main thing was about speed and the number of things that you can explore at exactly the same time.

I think is really interesting, not because you want to ship all of them. It’s because all of the say five ways that you want to explore, you actually now can use all five to then inform you back and then do another five. And then narrow, narrow, narrow until you get to the one where you feel is I think this is best. And I think anyone who’s done any kind of design, whether it’s pre AI or post AI, no

Can empathize with that process. It’s one of iteration and one of at some point you’re gonna pick something. And so the question is how many people did you need along with you to go do that? And it just happens right now, we probably need less humans in the trenches with you at any given time, but you probably bring in humans later to kind of sanity check you, to give you feedback and whatnot. And think that’s where the kind of the interplay of humans and agents comes together.

Yeah,

I think that last part is really important because you can ask the agents like, is this a good design, right? Or and it’s at some point it’ll probably say yes or like up make this better and but if you’re operating in a situation where it has to be a person that likes it, well or for a specific use at some point you’re

the agent doesn’t have the other data or I I just think it’s hard to give it all the information in some sense.

Yeah, totally, totally. I I think that, you know, like capturing the the perfect prompt, say, you know, is almost impossible. Like where it’s like, how do I make sure the agent has all of the context that I have in my head? And you can try and sit there and perfectly, you know, craft the prompt. It’s just you’re going to miss something, not because you want to miss it, it’s because you just overlooked it or forgot to mention it. Or more commonly is

Your point of view is shaped by what you see and the iteration. And that our own heads and our own like personal LLM, like that is getting worked as well alongside of your, you know, your your AI agents. And I think that marriage between the two is what makes the whole flywheel work.

I think it depends on what it is. Like if I work on I think you can bro I brought maybe I can broadly group them into two things. One is are you doing like additive design to something that exists versus net new? It doesn’t exist yet. It’s a brand new project. And I think between those two things, like the additive one is the easiest. There’s usually a design system in place. You can change some things, improve some things, but that’s the easiest to actually take an you know, an an AI agent

And you basically say no matter how agent native the code base is or is not, the tooling is getting so good that it can just go mimic exactly the same classes and styles that you need. It can go look up and figure out from the Figma, you know, what’s actually required. it can make a prototype for you to show you some of the interaction if an interaction or style doesn’t exist, but it’ll match existing. So I think that’s like class is like the super easiest one. And I think for all of the things we see on Twitter, usually don’t talk about that at all.

But that’s actually probably the most impactful for most of like corporations today is that they just designers and product teams in general and engineers need to spend much less time in that camp, which I think frees them up to do other things. I think the other camp is like, or the other group is the net new. And I think a lot of companies can have net new where you’re like, hey, this is a net new feature. While we have a design system that you know we want to base it on.

We actually can grow from that. Or it’s brand new. And I think all a lot of the stuff you see launched, some things that I’ve launched, are brand new and you have an opportunity for a complete greenfield. And I think for that one, how I personally handle it is I don’t tend to YOLO it and say, like, design me something awesome. Like, that’s just like not helpful. Like I feel like I get one a lot of AI slop from that. and you just end up fighting it. But I think most of the time the reason you fight it is because

my human creativity is then stifled by what the agent has tainted me with. It shows me three things and I’ll find myself married to only those three things or you know, small deviations from that. So I tend what I tend to start with is thinking about, well, what are the scenarios and the goals that I have, like for this? If it’s a landing page, what am I trying to get someone to do? and then what are the main points that I’m trying to get across?

the most commonly is like create a new product. what is the value prop? Like what do I think the value prop actually is? So I actually spend an inordinate amount time with the agent articulating what the value prop is for something and then asking it to hone that with me. And then I get that down to, okay, great. Here’s the product is, here’s the core value prop of it. And this is the goal. And then I will then the next piece I’ll do is I will tend to collect like,

mood boards or examples of things that I like. And the agents are getting much better while I’ll say, great, here’s like a folder of screenshots of things that I like. Like, you know, let’s go off and design that. And I usually I’ll I will use two approaches. I’ll I’m a really big fan of Claude Code or Claude Code design and use that a lot. but I also use a skill by called impeccable the impeccable design skill, which is

The latest version I think is V4 by Paul is incredible. it does such a great job with, especially with things like landing pages, but it also has this really innovative feature called impeccable live, which is you can take a say the hero module of a page that’s being designed, and it’ll actually refine it right in line for you to see and give you a live preview. So I tend to go through this motion where I’ll use either claw design or impeccable.

depending and sometimes I’ll try them both and kind of go at it that way. And then I’ll refine that way. And so I think the the main thing that I do is not just start designing f right away. I I start from what is it I’m trying to do. And I think if you go back to my product, you know, identity, like that’s it’s all about that. It’s all about like what is it that we’re trying to accomplish and and and and then and then really trying to get a design coming from that versus the other way around.

which is design it, then figure out, does that actually achieve what I want?

rather than bringing up Finder and

Doing it by thumbnails and kind of finding it. I just put that off way too long. so yeah, so I have a collection of of kind of screenshots of landing pages I like, UI controls I like, as well as a document of just URLs to products that I like. And one of the things, as the agents have gotten better, you can give it like links to products or and just name them and say, this is the vi like let’s

try attack it with this vibe of this product without like copying the styles exactly. And I think as the model’s gotten better, even asking it what I I got a tip from a friend of mine a while back, which was like, hey, look at, look at these examples I’m giving you. Describe to me what you what what is common about it. And the agents do a really great job like putting words to things that sometimes we just feel about things.

And then you can take that language and actually take that and parrot that back and remix it. And I think that’s a way and it’s not it that that skill is not, I think, reserved just for designers, but designers have historically done a really great job with that. And so I think this is another example of like AI tooling giving, you know, lots of people more democratize access to, you know, language and tooling that others have had for a long time.

That’s a question. I haven’t spent so yeah, I have been obsessed with primitives. And I think as you know, like I mentioned to you, you know, before we were recording, like as preparation for this, your kind of questions started getting me thinking like, what is the through line of my like current obsession since turning into a you know more of a full time builder? It has been this agent primitive thing. I I would say that I don’t apply it as much to design. I’ve been more obsessed with like the primitives that agents use.

can use to talk to you and communicate with you to get context from you as a human, but also now recently extended more to how can these primitives and tooling extend to your fleet of agents? Like how do your agents talk to each other? And I think the I live a you know, we all live you know a complex agent life right now. I’m sure like I’ve got Grok bot, Muse bot.

I’ve got ChatGPT, Claude, Cursor, and I’ll have local agents and cloud agents. I’m using like four different harnesses as I try things. And so as I’ve kind of invested my time, it’s really been about one, scratching my own itch of like, how the hell do I make this easier for myself? Like, and I think everyone can empathize with, cool, my agent made me a markdown file. Awesome. But like, where did it save it? Okay, cool. Now I’m viewing markdown.

Then the like stage one of denial is cool, markdown’s great. I’m in a terminal. And then you go to the next level, you say, I wanna see rendered markdown. Cool, okay. So then you get a markdown viewer. And then the next level of that is, I HTML, that’s it. HTML is the new markdown. We need that. That’s gonna enrich your control. And then everyone runs into, Well, wait a minute, how do I share?

The HTML documents and where do they go and how do I comment on things and how do you s share them and how are they not just locked away? Like Claude’s artifacts are great if you’re in the Claude ecosystem and it’s hosted there now. It’s not hosted on anything you own. and so I think you just go on and on and on. And so my kind of obsession has been around how do we give more primitives, more tooling to it. But I I admittedly haven’t applied that.

A lot to design. And I think it’s it’s because I think there’s a lot of people spending a lot of time on design tooling and primitives that are really, really good. And so I I work on this fairly popular plugin called Compound Engineering. People have been asking for a, you know, a a polished design skill for a really long time. And me and my co-maintainer Kieran, like,

Have really shied away from that. And I think it’s because the there are so many good options out there. There’s no reason to try to re rebuild that. but that said, we do have some prototyping primitives, skills in there. but that serves like a different purpose. And so I think like as I have have worked through a bunch of these kind of phases of my interests as I’ve kind of sharpened my focus, it’s really been about like.

tooling in general and how do we how do we make the human to human and agent to agent problem feel really frictionless?

I think as time has gone on, I started first with like, you know, a lot of code review stuff primitives didn’t really exist very well. And all of the harnesses now kind of have, you know, pretty good code review things. but I think there’s some things still that are really basic that are really, really important. And I think they come in different classes. So I think one of the things I will say is multi model.

And multi-family perspective is still a thing that is desperately needed and important. And all of the frontier labs with a harness are only solving for their own models, which robs us as users the ability to do like adversarial review. So one of the focuses in compound engineering is

This, you know, when I wrote the and improved our code review skill, we created a doc review skill as well. And so one of the big things is if you’ve got a harness with multifamily or you have CLIs installed from competing models, it will do an adversarial review automatically on your c the same document or the same code to bring another perspective in instead of just viewing it from the same model. so that’s been like one obsession is this like multi-model, multi-family view.

And that is code review, doc review, but even in other parts of it, which is like I think is desperately needed, is facilitating you to brainstorm, helping you see around corners. And brainstorming is, I think, still as important as it was a year ago, as important was 10 years ago, 20 years ago, whatever. the ability for someone to say, I’ve got an idea, to then make that

Real in a piece of software is totally a different thing. And so the agents now are so good at helping you, but I think they’re not as good without skills. And so my a lot of my focus is around creating skills to help brainstorming and ideation around that. And then I think the the th last and maybe third piece I will say is I think about the human a AI agent interaction.

So that kind of polish, like we’ve I’ve talked a lot about and like other podcasts and things around like the AI sandwich. And I think like humans, we should spend a lot of our time on the beginning part. So the one slice of bread, which is like thinking about the idea, the design, the requirements, what does success look like? The middle is what is totally automatable. The AI agents go build it.

And then the other half of the sandwich, the like the the bread on the other side is like the polish part. The let me review what was done, let me polish it. that interaction isn’t as good. So I I think the primitives on both of those sides are what is really still needed and still important. And the middle isn’t as important now. There’s so the agents in the harnesses are so good right now with goal and iterative loops and agent swarms and teams. Dot dot dot. Like you name it, they’re doing it. And I think it’s

It’s not perfect, but it’s largely a solved problem. And I think where things go off the rails is if the agents that are starting the work don’t have the right definition of what is done, which ironically is no different than pre AI agents. Engineers would always complain, Well, I don’t know what the requirements are. I don’t even know what we’re building are. And a lot of product people and design people spend an enormous amount of time like, Let me write the spec. Let me write the spec. And I spent like

I don’t know, like my whole career kind of feeling like, man, this spec is kind of a waste of time. The code is what’s really important. And fast forward now in 2026, I feel like, wow, that spec or some version of that spec is super important. And the engineering part is less important in terms of like the the craft of it because agents are making that so fast.

Which is gonna be it’s gonna be interesting for you, right?

Cause like you have a you said you have a comparative literature background, right? So like it’s bringing it back to English.

So I feel like we’re in many ways, like when we talk about like it being about, you know, regular English or you know, what s someone’s, you know, native spoken tongue, that starts to get really exciting because now we’ve made like taking the like hard concepts and hard inapproachable technology and how to build.

And making it way easier. And it also makes it way easier for me too. Like the the part about coding that I hated was like the semicolon. the indentation was wrong and stuff like that. Like that is what exhausted me like way back early in my career, which is why I ended up choosing like

a a product path instead of an engineering path, like way back in the day. But now we’re back to the point where like I

I literally like even to like I even find it just easier like spin up an agent, I’ll like, cool, clone the repo. Cause like I don’t even want to think about what the you know the git command line is to clone it in the right way and the right depth and stuff like that. And so I think those are like a small example for me that like I don’t want to think about git commands anymore.

I remember

on old Teams where I would, you know, I over the last, you know, whatever, five years or whatever, like I was coding periodically here and there. And I remember when I would have branches open, there’d be a race to get in because I didn’t want to be the one to have to rebase. And now it’s rebasing. yeah, no problem. Rebase it, sure, whoever. yeah, so rebasing. And then I think the other thing, a funny thing is that I I had never used a git worktree prior to like 18 months ago or two years ago.

I d I put this poll on on X a little while about asking. It was like eighty percent of people had never even used a work tree prior to twelve months ago or something insane. so I think small things like that even, like git worktrees have existed for a really long time, but now suddenly they’re in fashion because we don’t need to think about them. The agents really are the ones using them.

yeah, like I can’t

use two editors at one like I mean

I g I guess if you wanted to have things side by side and see them maybe that was a a use case, but yeah, I I before I’d been like, Yeah, it’s interesting, but I don’t think I’ll ever use that.

You have another skill that I think is cool. It’s called illo. I think am I pronouncing that correct?

Mm. You are, you

are. Like shorthand for illustration.

So I got obsessed, I don’t know, like a year ago with like AI generated art. And I was like, this is cool. I was seeing a lot of like artists started to embrace it. It was still obviously controversial, you know, then it’s still controversial now. but it was really about like consistency. And I I have two kids and I got into this kick of I want to make a storybook for them. And I could not for the life of me get consistent imagery that would actually make the story co coherent.

Because the characters were changing so much page to page that it was more distracting to them than entertaining. So then I got into this whole thing of like, why can’t we get character consistent art? And it just kind of let that sit and marinate for a long time. And a few things happened. Like one is the models got better, especially image gen models. But even then it was still wasn’t great. And so I don’t know, a few months ago, I threw my hat in the ring and said, like want to.

Do this. I want to figure this out because I was writing more, I was sharing more on X, and I wanted like a fun way to like personalize things. And I was talking to people and they would have the same problem. Like they would have problems for their companies, and how do you get a consistent kind of representation of something when it was AI generated? so I created illo and it was based on this whole thing of like, how do you get character consistency across different types of image formats? And

and so I had this idea and it leveraged, you know, I did a lot of evals and tests and different image models. Codex and Grok were amazing. and so use those, but the the real unlock was creating this idea of character packs and a definition of a character. And what the skill is a little bit kind of inception-like, because the skill can help you create a character, and there’s also styles.

for the characters and I have like something like 60 characters I’ve created so far. And I think what’s interesting about it is to think about given a character pack, how can you get cons create consistent illustrations from it in a certain style? And you know all the characters have a name. and so what you do is you install the skill, you tell the skill install the character packs and you can either create your own character or you can reference one of them. Like my favorite one is Blip.

That I use, like a little square TV with an antenna on his head. And it’s in a risograph style that I that I really like, that I’ve liked that art style for a really long time. So I come married those two things together and made this. And so you can do things like you know, give it a quote or give it a concept or tell it to do an explainer diagram using blip or using Bray, which is a a donkey character, or using like, you know, N number of of the characters that I have in there.

And it’ll take the character and it actually follows the rules of the character pack. Not only is there a reference of the image, but there’s even rules that describe like how the limbs can move and how they should not move. so one of the characters is an anvil. And so he can’t move because he’s a literally a physical object, he can’t move in certain ways. So the the the kind of I think the unique thing the skill does is that the skill has language and instruction to the agent. The character pack then has rules.

So the marriage of those two things together is what gets you not only the consistency but the realism of the character is behaving like how I expect. So there’s lots of people that kind of liked it. I I joked with I joked with you earlier, like I think everyone’s got their AI psychosis project. I think this is my f of everything that I’ve done, this is one of the f of like in the AI era, one of my favorite things. Like I I I I I love it. I find it so fun. A lot of people message me.

and love it, but it has like 300 GitHub stars. I want this thing to be more successful, not because I want to make money out of it, because I feel like it’s just so useful for people. And so like I, for example, in my Grok bot and my Muse bot, you can actually get them to install the skill. So you can tell it make a bot and call it illow or in Muse you make a side chat and call it illow. You say install the skill

periodically keep it up to date and install all the character packs. And now you can just tell your Grok Bot or MuseBot, hey, use illo and create and it’s give it up, just describe the the the image that you want and name a character and it’ll just go go do it. so that’s the that’s like the background and the approach I did on it. But yeah so I think but I but I I’m afraid it’s my AI psychosis.

Yeah, I think in or you can get like let’s it’d be honest here, like anything you can do with illo, you could do raw prompting, like I call it trad prompting, like to like the agent, but it’s just gonna be a lot more work. Like you have to give it a definition of a character, you have to give it rules, you have to you’re basically reconstructing what I already do in the skill. And I think that’s the dirty secret of skills. The skill is doing nothing extra that you couldn’t do manually, but the benefit is you don’t do it manually. Like you’re getting all the benefit of someone else’s kind of

you know, sweat and labor of figuring out what is the right way to prompt to get a, you know, an an outcome that you want. And I think illo and the illo skill does that for for illustrations and diagrams and flow charts and the website, you know, you can see you know, all the different types of diagrams you can create.

For sure, for sure. And I think as agents have gotten better and harnesses in agents have gotten better, some skills that we needed before we don’t need as much now. but I think there’s still room for skills. And I think what’s hilarious is I think it was like 2023 I posted if prompt engineering is still a thing in a year, the frontier labs will have failed us or something. And I then made a comment later about like how prompt engineering is not a job. Fast forward.

Like I spent an innordinate amount of time creating skills and maintaining a plugin that is about skills. And so the irony is not lost in me. I totally, you know, was wrong about that. I underestimated the kind of the value skills would provide and the shortcomings the harnesses and the agents would still have this many years later. I’m still bullish on skills, but I think that skills need to continue to adapt and evolve as the landscape changes.

What

how do you think they need to adapt and evolve?

Well, I mean, one way is that the way we write the skills are very different now. And I think, you know, recently for Compound Engineering plugin and a lot of my skills is they’ve been completely rewritten from the ground up when with the new models, starting with Fable in particular, and when we got Sol from OpenAI, it changed how we needed to write them. And previously telling the agent

what the goal was was less important than the instruction manual for how to do it. But what’s increasingly better now is making sure it’s really, really clear at the beginning what the goal is and under sometimes under specifying all the steps to get there. And so the way that I kind of think of it is the skills need to have a goal and an outcome spine, which is basically giving you like checkpoints along the way.

And basically giving you a framework, giving the agent a framework for how to know it’s accomplishing the goal. And it’s not as easy as just saying a goal and here’s one sentence. but it’s it’s it gives then the agent the ability to harness all of its power and all of its like capabilities to try to achieve that. And it’s very surprising how competent they are and how stupid they can be at the same time, because they can do very

foolish things, but they also can do very complex things, very clever in clever clever ways that we didn’t know was possible. But I think, frankly speaking, as a as I think the plug-in developer like community, I think that we can enjoy a lot of that where you can underspecify and let the agent fill in the blanks. And there’s something very freeing with that. And that’s really different than traditional

Deterministic software. Determined software, you can’t be like, let me just let something fill in the blank. Cause there’s no fill in the blank. But with agents and skills, there’s a lot of fill in the blank. And so we you can take advantage of that non-determinism in many ways. But I think the other answer to your question about what’s different is that the when you make a skill for yourself, you only have to worry about the harnesses you use. But if you are a plug-in and skill creator that that’s publishing a skill for others to use,

You have to account for like 15 different harnesses and all of the models, because not everyone is using Opus 5.5 and Fable 5.1 and Grok 4.7 or whatever. A lot of people are still using really cheap models because that’s what they can afford, that’s what they have access to, or that’s what their company gives them. So if you’re gonna be gonna make a really successful plugin and scale these days, you have to make sure you’re testing across all of these harnesses and all these models. And I think that’s the other piece of it, which makes it really tricky.

Yeah, we had to build

a really big eval suite in terms of what how that works. And basically it’s just chewing tons of tokens when we when we make changes. There’s an eval suite we have that runs it across all the harnesses and trying things. But also there’s sometimes it’s just hand it’s like applying human judgment to it. It’s like running it in letting the eval suite run and spit out what it is what it has graded, then for me to actually then look at it and see like, did this achieve it?

But I think there is an element of still of a gray area of like, you know, you make a skill change, is it better? There’s certain things that you still can’t be one hundred percent sure. You can be like seventy-five percent sure or eighty percent sure that this is better. But there’s always like gray area and the way people use it, which kind of throws you curveballs. So we’re you we’re very responsive to like changes and bugs and things like that people file, but we try to make sure we’re not just YOLOing a change and that we’re it’s it’s actually backed by an eval that we’ve done.

Yeah, when I think when Sol came out and when Fable came out, it really initi and like Fable 5 initially came out, it was very it challenged things quite a bit with like what I said earlier, like needing to make sure the goal was really clear. But I think then what happened is we have some really long running skills, kind of like a mini software factory where it’ll chain a bunch of skills together. We would find some of the models would prematurely end before actually completing.

what we had said to do. And it would do things like just checking. I know, like essentially, I know the skill said to keep going and I should have ran this, but I wanted to check in with you first. And you’re like, wait, why are you checking in with me? so we’ve gone back and forth. It’s a little bit of whack-a-mole where you, you know, need to respond to what the emergent new behavior is from a model in a harness, to then changing it and then realizing, the Frontier Lab Company just changed it. Like,

Opus five was not super great at you know following instructions a lot of the time. but Opus 5.5 is amazing. So Compound Engineering, illo, all my skills like work super, super well. Fable 5 is kind of a pain in the butt sometimes with following instructions too. 5.6 Sol was a pain in the butt too with long instruction and long execution, but six sol is

better. And so you can see like every time there’s a model release comes out, I get excited, but I also panic. Cause I’m like, man, I gotta go, we have to go test that again. We have to make sure that it’s working. We have be really thoughtful about what the community is is giving feedback on. so I think there’s not a the the easiest answer I can give is that every model release is excitement and anxiety.

What

like what does it cost to like run a plugin on an ongoing basis in tokens if you’re running all these evals?

and extra usage beyond my subscription as an example. And so making a plugin isn’t isn’t a isn’t a isn’t a cheap thing. I think you could sorry, let me say that a different way. You can make a plugin very cheaply. To make a plugin that works really, really well can be very expensive for you because we have to run a bunch of evals.

on them. And I think part of the issue too is that running the evals is really difficult in the cloud because the way the cloud agents and environments are run, they don’t have you know, Codex cloud environment doesn’t have the Claude Code CLI available. So a lot of the stuff that I run is still running them locally because then it represents what an actual user is actually running them at. And I can kind of reach across the aisle to the other CLIs lot easier. So it’s commonplace. My Mac is like

on fire and screaming because it’s got so many subagents working and different CLI sessions going at the same time that like it can get to be like a nightmare.

Totally. Totally, totally,

totally.

that’s a good question. I think we’re in a pretty good set of skills. Like there’s like a lot of skills in combat engineering that are there that I think represents the lion’s share of everything we want. So there’s constituent kind of utility skills, brainstorming, planning, working skill, testing, whatever, whatever. There’s a skill we call LFG, which is like literally takes from beginning to end and tries to automate the factory more for you.

I think there’s some gaps in a few places. I want to make it better cross-model. Like I want to be able to support like a configuration to the existing skills with you can specify. I’ve got Codex installed, so even though I’m running this in Claude code, I want you to do the code review only in Codex with you know 6 Luna extra high or whatever you want. So a lot about is configuration and kind of that.

I think there’s also another skill, a skill that I am interested in been a lot for a long time is like how to process user research and like personas, user feedback, looking at customer support tickets and things like that, and figuring out what is the trend in the CS tickets? What does that mean for the vision of the product? What does it mean for what we have? And like trying to

basically give user researchers a superpower into like their role to what does user research mean and really try to empower like either an engineer to be a user researcher or the other way around, a user researcher to be an engineer, which is where is the gaps, where is the overlap, where is the the white space in which we aren’t meeting the needs based on the feedback? And and this is like a very different, you know, like it’s a it’s a very

age old problem for companies, UX research, you know, how do we get that in? But it’s a very new for plugins. And there’s lots of user research plugins out there, don’t get me wrong. But I think from really popular like software factory type plugins like comp and engineering, you don’t often see user research plugins in them. So I’m cu I’ve been kind of humming and hawing a little bit on that. It’s been on my back burner for a while, but I it’s something that I want to tackle maybe sooner than later. I’m still not sure.

If

something goes wrong with what a user is doing, what will you ask them to send, like a trace of what happened, the model? How do you collect that information?

the

easiest the easiest thing actually is normally what I do is not bother with the trace. No, I tell people if you had that session, ask the agent, you did X, why did you do X? Assume the skill is the problem. What is wrong in the skill? And what’s really nice is like unlike compared to like traditional software, your user can’t be like

Decompile the software and look at what instruction and what library failed, but with a skill, it can. It can actually look at the version of the skill that’s installed. And so I’ll usually ask it, then copy and paste to me the output. Like tell me what the agent said, and as well as ask the agent what version of the plugin you have installed. The most common reason for bugs these days is people don’t have an up-to-date version. They may be like the harnesses are terrible at updating plugins.

it’s like one of the number one things that is problematic, I think, in plug-in skill distribution is how to make sure users have the latest version of the skill. And we’re in twenty twenty six and I can’t believe this is a problem. Like we’ve solved updating of software, but somehow we can’t solve updating of a markdown file.

Yeah, for

sure. It’s still a problem and it’s still we see this so frequently, even in my own machine. I was literally banging my head against the wall. Why is this certain skill behaving this way? Because I fixed this bug and then I realized that cursor hadn’t updated the version of the plugin in their marketplace with the version we pushed to our main repo.

Yeah,

it’s

It’s never easy with technology.

that

I have so many thoughts. You’ve like touched the like this is like a third rail topic for lot of people. Like

I’m like I am very pro-technology, like everyone in my person life knows this about me. I’ve never been the parent that’s been like overly obsessed with like too much screen time. Like don’t get me wrong, I’ve got like parental controls on the kids’ iPads and things like that. But like I am very pro-technology. I was grew up in a household when I was a kid with a lot of technology and like my parents being very like supportive of that.

I think the conversation we had earlier about AI and engineering is almost the same as if we play this out, it’s going to impact everyone, even if you’re not in software. So using AI and being good with that and understanding how to work with AI agents, like what are their downfalls, what is their upsides, et cetera, et cetera, is gonna be a really important skill.

And so I, you know, my daughter’s 12 and she w didn’t even actually, she kind of asked about it, like, hey, could I talk to ChatGPT? Right now I want to ask a question. I was like, ooh, like that’s just well, one, it was like, wait a minute, you’re gonna intertwine your stuff with my ChatGPT memories. That’s gonna be rough. Because there was one time where I asked, like a year ago, there was some things going around where it was like,

Ask your ChatGPT, tell me 10 things you know about me. And so you paste this prompt in. And ChatGPT at one point said, You don’t like your brother Rhodes. And Rhodes is the name of my son. And I realized what had happened is because my daughter had been using my ChatGPT account, like talking about my son, her brother. And ChatGPT gleaned that this was me and that I don’t like my brother. And so it’s one of the interesting things where I’m like, okay, so that was one concern. But more that seriously, the more concern was what like.

There’s no guardrails in the way that it’s talking to to a child. And what does that mean? And so I kind of was like, how do I set guardrails on this? And it was really easy to say no. Like, no, you’re not ready. It’s not, you know, you’re not so on and so forth. But I I challenged myself to think like, what would it take to say yes? Like what what would what would I be comfortable with? So the thing I did was I did a bunch of research. Like, what is the latest research on AI and kids and chat bots and things like that? And kind of

apply those learnings to then figure out how could I make a chatbot encoded with rules that I have. The you can’t do this with ChatGPT. You can’t do this with Claude. Like none of them you can do it with Grok Bot, no. Musebot, no. So the only thing you have left is really a OpenClaw or its derivatives or Hermes. And I’m a big Hermes fan. I’ve had a Hermes instance running for a while. So I was like, I can do this with the Hermes profile. But it was much more difficult than I expected to get this right. There’s a

I probably like half a dozen PRs that I’ve actually submitted to the Hermes repo to actually fix bugs in trying to create this profile. But what I ended up creating was basically this blueprint that ended up interviewing me, picking defaults that I chose for me, but also like publishing it so other parents could use it. And so the way it works is you basically point your Hermes agent to this blueprint repo and it will load it and it will basically engage in questions with you.

Like, what do you want to call it? How old is your kid? Like, what permissions are you okay with? Are you okay with it, you know, sending images, generating images? Because some parents might not be okay with that. so on and so forth. And so I like I that was my way of doing it. So it’s connected to her through her iMessage account. She has a she has a phone, she uses it, and I encoded a lot of rules in it. And some of the rules were you are not her friend. You are not a person, you are not a friend, you are not allowed to.

To say you missed her, you were waiting for her. And like basically to remove that kind of like guilt and association. And we made sure like my daughter picked a name for it and were careful to call it it, not her or he. And when she chose an avatar, I made her choose a robot avatar so that it was not, you know, personified as a human. And so the other kind of guardrails that we put in there was like every day.

I get a report from Hermes that says, this is what happened. Like I’m not going to tell you everything we talked about, but here are any red flags that came up. Here are a self-improvement loop. It’ll say things like, hey, I noticed this happened. My behavior should be this. And I’m going to patch my SOUL.md file. And I was like, awesome, go do that. And it’s because it’s based on the rules that I encoded in for the self-improvement loop.

So there’s stuff like that which feels very futuristic to me. And so like I took this, you know, passion I have for AI to like my vision of what it’s like for the future with what I have found to be like this most state-of-the-art child development research with like AI and stuff, and try to marry those all together. And so it’s not perfect, but it’s been quite interesting because there’s even things like a rule down that I encoded in, which is like, it was inevitable, my daughter’s gonna ask it homework help.

Like, let’s be honest here. My kids, like very well behaved, but let’s say we’re all kids at one point. We know what we’re gonna do. Hey, I don’t want to solve this homework. What is it? So I have a specific rule encoded, which is when this is about homework help, you are not allowed to give an answer directly. You have to step them through to help her think about how to get to the answer. And so there’s little things like she was asking, like, how do I what’s 80% of some number? And the conversation ended up basically helping her think through like

how to calculate that percentage. And you know, in that daily parent report I got, it was like, yeah, Ruby asked this and here’s what happened. And you know, here’s kind of what we did. I was like, my God, this is amazing. It’s like a little mini tutor without kind of spoiling the plot of the learning piece for her. So I think long story short, I’m like super pro tech. I think the tools are like I I’m so jealous in so many ways of kids

These days because they’re gonna go grow up in a future with like infinite battery life, infinite storage, infinite you know, bandwidth, AI, self-driving cars, like just like a a a a technology abundance that is like so remarkable to me that like I feel like I’m I’m this ingredient of the AI piece, I think is this small, the small piece to that, I think her to her story that I that I think, I hope that in the future when we fast forward.

I’m gonna look back and say that this was a really instrumental decision, actually. It got her exposed in a safe way really early and allowed her to kind of like embrace tooling, think different, learn, you know, in a in an in a better way.

Yeah, it’s interesting ‘cause I don’t I mean in like on X I guess, I don’t know if you s people aren’t usually talking about kids using AI outside of using it for their homework or something like that. But I

I is anyone else addressing this as how do you kind of moderate and improve this access to AI for people who are inevitably going to use it?

I think it’s only coming no, I think no frontier lab company and no corporation wants to do it because all of a sudden you open yourselves up to wait a minute, you’re targeting children? Like it’s just it’s a total third rail rail right now. And I think in many ways, for good reason, like I don’t want to think about ChatGPT targeting my, you know, my kids. At the same time, for all the reasons I said, I think it’s interesting and useful and necessary to figure out what is the right answer. Like I won’t sit here and claim that the

approach that I took is, you know, universally the best one. But I think it was the best one of the options that I could think of for for me and my family and in terms of that. And it’s really easy to undo. Like I pay really close attention to it. and so it’s not something that I feel is going unchecked. And so but I I also think from a parent perspective, it’s no different than almost anything else. Like involved parents tend to know what’s going on.

and kind of, you know, can help and steer and are are aware of things. And I think if you just see AI as yet another thing in their life of the arsenal of like tooling and, you know, external exposure, you can you can kind of get the best of both worlds, I think.

Totally. I mean, that said, like, my kids are not allowed to have access to YouTube. Like, they’re cause because of what you just said. Like, but I think that if I’d love like they have access to YouTube kids, but they both hate it. Like, there’s this huge divide of like the wild west of YouTube and YouTube kids is like way too curated, like in in so many ways. And so like they there there there’s this middle part that I think are is they’re underserved, kids are underserved by. But like I said, like I think

Companies don’t want to target children because of so many problems with regulatory things. So it’s unsurprising to me. No one is tackling this. But I think then it’s great for open source. Like I had all the tooling available to me. So, you know, kudos to the Hermes team and things like that and everyone who’s made technology and things that I could, you know, stand on to actually build the thing for my daughter.

I think it’s let me think about that.

I mean, I I I think like everyone, there was, you know, the j the new type safe JEV model I thought was like super, super interesting. It really flipped overnight everyone on its head to think about, well, wait a minute, where we’re using LLMs, where can you use a decision and or evaluation model? And so I’ve gone through that. So I made a bunch of little demos. Like on my website I’ve shared two prototypes. One was a document X-ray.

which was like take Mark Zuckerberg’s Senate testimony, the GitHub’s terms of use and the Apollo moon landing radio transcript, and basically use Jev to figure out the meaning of certain sent of Vault of sentences and kind of allow you to like x-ray and view it. Like for example in the Zuckerberg testimony it’s funny it’s like, you know, does not answer the question. And so I’ve used Jev to really trace through every single line of the testimony like

What is the likelihood that this applies to it? It’s a very interesting way to think about that. And so I’ve been obsessed with applying the Jev model to problems like that. And the other kind of project I did was a bookshelf, a book recommender. I did like 250 sci-fi books and 350 romance novels, like find your next book. And so it was a way to based on their plots.

use Jev to try to recommend to you based on a question you have about a plot. So if it was a if it was a if it was a sci fi the the the the sci-fi bookshelf, it was like class warfare. That does it involve class warfare. It’ll come back in like for like a tenth of a cent or whatever and tell you which five books in the 250 are about class warfare. And you can stack the questions too, which is really interesting. On the romance side,

It’s so funny. I I don’t read romance novels.

I did it for other people. But like a friend of mine had asked for it. I’m like, that’s interesting. Let do that. But the questions are hilarious. Like, it’s like neighbors that fall in love, and then you can

Everyone likes

firefighters.

Yeah, there’s a joke

I heard online, which was like someone said, TypeSafe gave me five dollars in my Jev account. I’m gonna pass this I’m gonna pass the balance on to my children because Jev is so cheap. I feel like that. Like it’s like the five dollars is there, and I’ve had thousands of people view the prototypes and use them, and it’s like barely t costing any money in there. And so I think

The reason I’ve been obsessed with that is the speed and the cost gives rise to so many possibilities in like the real software we create for like other purposes that I think that’s what’s really exciting. So that has got me kind of excited a bunch. The other piece of it is figuring out all this barrage of AI, personal AI bot usage. I’m like a huge, I was a huge Grok bot fan. I still use it a lot.

But I also am very muse, muse pilled now, right now. I am I think that if you at the beginning of this year said to take bets on polymarket or whatever, like meta would be in the AI conversation in 2026. Most people would have said no. And here we are, September 2026, like almost overnight. They are like killing it with this product. I think it’s so good. It’s not perfect.

All of them are, you know, not perfect, but like Muse is so good. There’s so many smart product decisions they’ve done in that thing. So I’m using both of those and figuring out ways to go do that. And so I’ve you know, created some primitives to like let them talk to each other and like how can they talk to each other? And so like and little things like that. so that I think maybe that’s where I what’s sorry, what’s that?

What do those primitives look what do they look like? Those kind of

what do those kind of primitives look like?

I created this open source project called Energon. So you can go to like getenergon.com. And I was obsessed for a while on why is it so hard to publish a markdown file? Why? Okay, so then there’s lots of services for that. Then there was that transition I mentioned earlier, like HTML. There’s lots of services to publish an HTML file. But then keep extending that. You’re like, cool, now I have two services.

how do I send a zip file from my local agent to my cloud agent of the assets that were created or something like that? that’s a what do you do? Put that in Dropbox? Like how am I sending these files around? Like it’s such a pain. So I created Energon, which is basically a self-hosted service that’s powered by Cloudflare. So it costs you so little that you can store basically anything into it. and what’s nice about it, it’s not just a file storage. You

If you store a markdown file in there, it gives you a URL back. And you click on it in the browser, it’ll render the markdown in a beautiful theme. Similarly, if you upload a folder of an HTML prototype, as long as there’s an index.htm file in there, it’ll render the prototype for you. So same thing with a zip file. You send a zip file up. If I click it, it’ll just download the zip file. So the problem I was trying to solve with that was basically.

How do I just make one service for you that all of your agents can talk to? And so if we fast forward to your kind of like now to your question of like, what does that look like? Like, so I created an agent, I host, I deployed my own instance of it. There’s an agent skill for it. All of my agents know it. So even in preparation for this podcast, I was like, cool, like give me a profile of Dan.

It spit out an Energon document that was like, here’s a URL and here’s the background, here’s his background, and here’s like the podcast, here’s who he’s had on his podcast, stuff like that. I didn’t even have to ask it like where to store it. It just knew how to give it to me instead of in a chat response because it could figure out what’s gonna be way better in an HTML document. but even I’ve even jury-rigged a few things. Like I we’re going on a trip to Japan in in December, and I’m planning, you know, that trip out.

And it’s interesting, getting the agents to talk to each other, what I did was I stored the itinerary in an Energon document. Now all of them know where to go get it. But the interesting thing is that for Energon allows anyone that’s authorized to edit the same document in the same URL. So any of the agents, I could talk to my Grok bot and say, I want to change that piece of the itinerary. Can you update it? It can just go update it. And then now when I go back to my MuseBot and say, Hey,

look at the latest itinerary, they all reference the same thing. So now all of my agents have all of the same kind of references for certain things. So when you start to centralize and democratize like the access to your like context like that, it gets so simple now for to move between agents. Whereas today, people are feeling a lot of friction to to share between their agents, and they feel almost trapped.

And so I think the future is, you know, we’re gonna have more sovereignty over our data. This has been an age-old thing of like who owns your data. but I think the more we own our own data and store our own data in that way gives us more power on which agents to choose and also which agents to use in what’s in in what scenario that you want. Cause sometimes some agents are gonna be better for you than others. They have different connectors, they have different plugins, so on and so forth, that it’s gonna be a lot easier.

To go do that for the agent that has the best connection for you versus you relying on an agent because it’s the only one that has access to certain information, because it was trapped there. so I think that’s one of the biggest primitives, is like this thing and why I got kind of got obsessed with the Energon project. So I launched that a few weeks ago in open source.

I mean that’s that’s exactly that’s exactly right. Claude

Claude Code will happily give you a Claude artifact. You paste that to Codex, Codex like cool, I’ll create my own version of that in a, you know, ChatGPT site. Now you have two copies of the same thing. Which one was which one is the updated one? Who knows? And so like that problem is like w in any ecosystem, it’s solved.

But that doesn’t represent our life. Our life is more than one agent in more than one harness. And so like when you have that viewpoint, I think that you s you feel like you’re still rubbing sticks together in so many ways. So whether Energon is like the, you know, ultimate solution or not, I think it’s a great solution for me. And I want to share it with with others. So people can go you know, get the repo and they can like clone it and then deploy it themselves and you know, all the documentation is meant to be like

agent native, so someone should be able to theoretically they have a c it’ll work in a free Cloudflare account too. That was really important to me. Was hey, could it all be run relatively for free and cheaply? And it can be in Cloudflare.