The argument in one line.
When prompts become the real source of truth and new model generations let you discard and regenerate the code, the craft of software development shifts from writing code to describing what you want with precision.
Read if. Skip if.
READ IF YOU ARE…
- You are already using Claude Code or Codex daily and want to see how a founder who builds AI dev tooling actually structures his workflow.
- You are curious how a startup keeps a codebase from degrading when the vast majority of code is AI-written.
- You want concrete takes on when Claude Opus outperforms Codex, and why a GUI beats the terminal for agent work.
- You are thinking about running 3 to 10 agents in parallel and want to understand the UX decisions behind a tool designed for exactly that.
- You manage a small engineering team and are curious what a fully AI-native development process actually looks like day-to-day.
SKIP IF…
- You are still getting started with AI-assisted coding — the workflow assumes multi-agent orchestration as a baseline.
- You want a step-by-step tutorial — this is a philosophy-of-work interview with demos, not a how-to guide.
The full version, fast.
Conductor co-founder runs every task through his own product — an agent orchestrator where each workspace maps to a Git worktree and becomes a PR. His operating rhythm is voice-first: a twenty-dollar gooseneck mic, dictate a task, watch the sidebar fill with parallel workspaces, review with keyboard shortcuts, comment GitHub-style, merge. His guardrail against AI code rot is slop-free zones — human-only codebase sections with literal comments saying the AI may not touch them. His deepest take: the prompts are the real source of truth now, because when the next model generation ships you just rerun them, and the old code did not matter.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:22
01 · Intro + desk setup
Charlie introduces himself and Conductor. Split ergo keyboard, gooseneck mic, monitor running Conductor.
00:23 – 00:50
02 · Talking to computers more
The twenty-dollar gooseneck mic as the unlock for voice-first coding in an open office.
00:51 – 02:19
03 · Using Conductor to build Conductor
Live demo: cmd+N for new workspace, voice task input, sidebar tracking, cmd+shift+Y to review, GitHub-style inline comments.
02:20 – 02:39
04 · Conductor on the go
Mobile demo: dictate a feature request on iPhone, desktop Conductor starts working immediately.
02:40 – 03:51
05 · Does Charlie still write code?
No. Caveman mode for rare file edits. Voice or highlight plus tell AI for small changes. Archive on merge.
03:52 – 04:17
06 · Feeling like the CEO of a little company
Dashboard concept: one view showing all agents, digestible reports, redirect or merge from one place.
04:18 – 05:47
07 · Other apps + customization
Spokenly (local Parakeet STT), CLAUDE.md plus skills files, always fast mode, Context7 MCP, dangerously-accept-all-permissions as default.
05:48 – 06:32
08 · Slop-free zones
Human-only codebase sections with literal AI-exclusion comments. Guards against bad-code feedback loops.
06:33 – 07:05
09 · Conductor tech stack
Tauri app, native Safari renderer, Rust backend, ~95% TypeScript on desktop, Elixir/Phoenix for web.
07:06 – 08:30
10 · Don't let the AI be your architect
Humans decided the workspace abstraction, the three-panel UI layout, the open-end button. AI loses the crafted feeling when it makes these calls.
08:31 – 08:58
11 · Where to give AI free reign
Human-written APIs and contracts form the core; AI gets free rein on everything built on top. Current boundaries are murky.
08:59 – 10:35
12 · Enforcing workflows + building conviction
Every workspace must create a PR. Conviction comes from dogfooding, not analytics or A/B tests.
10:36 – 11:16
13 · Codex vs Claude Code
Codex grinds through debug loops. Claude Opus is the creative partner for new features. Context-dependent.
11:17 – 12:00
14 · Why a terminal is not enough
Humans are spatial and visual. The GUI beat the terminal in the 1980s. Same argument applies to AI coding interfaces now.
12:01 – 12:52
15 · Thoughts on tokenmaxxing
Peak: 22K on tokens in July 2025. Always fast mode, always max effort. Keeps lines of code minimal to counter sprawl.
12:53 – 13:24
16 · Workflow changes vs 6 months ago
No more IDE for hard PRs. GitHub web app almost never used. Code review and checks live in Conductor.
13:25 – 14:16
17 · Most surprising Conductor use
A user built a mobile Conductor by spoofing IPC calls. Gary mode: all tool calls uncollapsed by default.
14:17 – 14:59
18 · What the world does not see yet
Human-AI collaboration frontiers: multiplayer chats, sub-agent communication. The conductor-of-an-orchestra metaphor.
15:00 – 15:47
19 · Code is becoming sawdust
Code is the byproduct of describing what you want. The prompt is the real deliverable. Rerun prompts on a new model.
15:48 – 16:35
20 · Malleable software + outro
Call of Duty modding as the model: crafted structure, personalizable workflows layered on top.
Lines worth screenshotting.
- Code is becoming sawdust — it is the byproduct of describing what you want, not the thing you are actually building.
- The prompts are your source of truth: when the next model generation ships, rerun them and get new code. The old code did not matter.
- AI can enter a vicious cycle — it sees bad code, then writes more bad code. Slop-free zones break that loop before it starts.
- Do not let the AI be your architect. Architectural decisions, UI layout, and core API contracts must come from humans.
- Running Claude with dangerously-accept-all-permissions is their default, not the exception — treating agent autonomy as a feature, not a risk.
- A twenty-dollar gooseneck microphone is the most important peripheral for AI-first coding in an open office.
- Twenty-two thousand dollars of token spend in a single month generated tens of thousands of lines of code — and they still try to keep lines of code minimal.
- Codex is the workhorse for grinding through a debug session; Claude Opus is the creative partner for thinking through a new feature.
- Feeling like the CEO of a little company — directing agents, reviewing reports, redirecting — is the interface metaphor that matters now.
- If you let the AI make your UI choices, you end up with something that does not feel crafted. Humans decide the interface; AI executes it.
- The terminal lost to the GUI in the 1980s because humans are spatial and visual. The same argument now applies to AI coding interfaces.
- Malleable software is the next frontier: software users can mod like a video game — crafted structure, personalizable workflows on top.
- CLAUDE.md and skills files are load-bearing engineering artifacts, not documentation — Conductor has hundreds of lines of hard-won context in theirs.
What running ten agents changes about how you think.
WHAT TO LEARN
When you stop writing code and start directing agents, the discipline is not about prompting harder — it is about keeping humans in the decisions that compound.
- Slop-free zones are not a workaround — they are architecture. Designate human-only sections of your codebase before AI touches anything load-bearing, or you will spend months cleaning up a feedback loop where poor code trains more poor code.
- The prompt is now the source of truth, not the file. When the next model generation ships, the only thing worth preserving is a clear description of what you wanted — the code itself can be regenerated.
- Do not delegate architectural decisions to the AI. UI layout, API contracts, and system abstractions need a human decision behind them — the crafted feeling users recognize comes from that layer of intent.
- Codex and Claude Opus are not interchangeable. Reach for Codex when you need to grind through a specific bug across many tool calls. Reach for Opus when exploring a new feature and you want a creative partner.
- Running AI coding agents at maximum effort with full permissions is a deliberate operational choice, not recklessness — it reflects treating the agent as a collaborator with real autonomy.
- A GUI beats the terminal for agent work for the same reason it beat it in the 1980s: humans are spatial creatures. Knowing where your chats, review panel, and agent statuses live in space is what makes multi-agent work feel manageable.
- Keeping lines of code minimal matters more, not less, when AI is writing most of it. A codebase that sprawls is harder for the next AI session to navigate — brevity is a forcing function for clarity.
- The next frontier in AI coding is not more capable single agents — it is the interface for directing multiple agents simultaneously, getting digestible summaries, and knowing when to zoom in versus when to wave the baton.
Terms worth knowing.
- Slop-free zone
- A section of a codebase designated as human-only — the AI is explicitly barred from contributing without every line being human-reviewed. Guards against the feedback loop where AI reads low-quality code and writes more of it.
- Workspace (Conductor)
- An abstraction around a Git worktree in Conductor — each represents one agent task, progresses through in-progress, in-review, and done states, and resolves into a pull request.
- Caveman mode
- A Conductor mode that re-enables direct keyboard editing of files — the name signals that typing your own code is now the exception, not the norm.
- Token-maxing
- Running AI models at maximum effort and speed settings to extract maximum output per session, as opposed to conservative or budget-conscious usage.
- Fast mode
- A Claude Code setting that enables maximum token output rate — not the default, but used here as the standard operating mode for all sessions.
- Gary mode
- A Conductor display mode where all AI tool calls are uncollapsed by default — built for a specific power user who wants full visibility into every agent action.
- Malleable software
- Software designed so users can layer in personal workflows and customizations — analogous to video game mods — while the core structure remains human-crafted and consistent.
- Context7 MCP
- A Model Context Protocol server that injects up-to-date library documentation into Claude context window, reducing hallucinated API signatures.
- Parakeet
- A local speech-to-text model run by the Spokenly app — used for voice dictation to Claude without sending audio to a cloud service.
Things they pointed at.
04:22toolSpokenly
05:09toolContext7 MCP
04:18toolTelegram
10:36toolCodex (OpenAI)
01:40toolLinear
Lines you could clip.
15:00
“Code is almost like sawdust now in that it used to be that code was the thing you were building. And now the code is almost just like sawdust that comes out of that process.”
07:06
“Dont let the AI be your architect.”
06:15
“We have like some lines in our code base that are like do not touch if you are an AI. Like this is for human eyes only.”
12:14
“I spent twenty-two thousand dollars on tokens that month.”
15:09
“Really what matters is your prompts. And when the next generation of models come out, you can just rerun your prompts again and the old code didnt really matter.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Hello. I am Charlie, the cofounder of Conductor, which is an app that lets you orchestrate a bunch of coding agents on your Mac. And we were YC Summer twenty four, uh, and I'd love to show you my setup.
00:22So a recent thing that I can't live without is this gooseneck microphone. $20 on Amazon.
00:28We are all trying to talk to our computers more. One issue with having, like, an open floor plan office is that can be pretty distracting. So one advantage of these is you can, like, lean over and whisper into cloud and be like, please merge p r three four seven five.
00:44And it's a a little bit less disruptive. I we we all got these in an attempt to encourage more talking to computers. I spend most of my day in Conductor.
00:52We're using Conductor to build Conductor. One thing that I do is I'm constantly kicking off new tasks. So I'm constantly going command n.
01:01That was actually a sneak a sneak peek of something we are working on, which is Cloud Workspaces. But I'll I'll I'll do command n, and then I'll speak into my computer and say, can you take a look at the latest linear issue and give me a a rough pass at how you'd solve it.
01:17Stuff like that. And then press enter. And then I can see that it's running in the sidebar.
01:21And while Cloud is working, I'll go to another chat. I'm very into keyboard shortcuts, so I try and make everything have a keyboard shortcut. So in this case, I'll do command shift y.
01:30I can see here that this workspace is ready to merge, so I'll take a look at it, give Claude a quick review. In this case, it's a pretty small PR. And so it looks good for to me.
01:39But quite often, Claude won't get things exactly right. And I'll give things a comment like a GitHub style comment and say, this looks a little bit weird to me.
01:50Why do we need this? Press enter, get cloud running, and then go back to a different workspace. A big part of how I use Conductor is like experimentation.
01:59Like, I'm always kicking off workspaces to try different ideas. Most of them don't make it in. So, like, you can see we have like, four PRs here that are in review, but, like, there's a bunch of random ideas that I've tried here that are in progress that, you know, may never see the light of day.
02:16If I like it, then it might get promoted to, like, an internal setting and then an experimental setting. Okay. So something I'm very excited about is on the go.
02:23Uh, I'm gonna just speak into my phone and say, let's add a new feature where I can change the theme to hacker mode. And then I'm gonna click conduct, And then my computer starts working on it and I can conduct on the go.
02:39You still write code today? No. Yeah.
02:43No. Very occasionally I will like edit tailwind classes or like open up an IDE to change like a dot ENV file.
02:51We actually added a mode that we call caveman mode which is you click this and you can actually type with your keyboard and like make changes in a file.
03:03Once in a while you do need to like make a change to a file by hand but it's like it's called caveman mode for a reason. Most of the time if I want small edits I'll like highlight and then tell the AI about my comments. Or I'll just like speak into my computer and say, that button looks a little too wide, like, can you can you make it smaller?
03:22By the way, this thing is now ready to merge. So I just wanted to show you. I can now click archive.
03:28And it's gone from my side panel and like merged into the code base. And this one I can see that there are checks running. And once it's finished, I can just like click merge and and get it in.
03:37We recently added this thing called status in the left. So when something like is kicked off, it's in progress. And then once there's a PR created, it's in review.
03:46And then once it's merged, it goes into the done folder. We have this new concept of a dashboard page where at from like one place, you can see what all your agents are working on and then like take them to the next action.
03:59But we're still messing around with like what the what the interface should look and feel like. But the ideal is like you should feel like the and you can see all your agents working for you and they'll bring you up like digestible reports and then you can put them in the right direction if they like need some correction or just merge it in if it looks good.
04:17What are your other main applications, main software that you use? I use Telegram a decent amount to talk to my OpenClaw.
04:25That's that's been a recent addition for me. I use Spokenly for text to speech. That's what comes up when I press control space.
04:32It's actually running a local model. It's running Parakeet. I have a really beefed up computer.
04:37So it's like a 128 gigabytes of RAM partly so I can, like, run local models like Parakeet. But as a side note, I have just recently ordered the MacBook Neo. Like, the the bottom of the line lowest RAM, lowest memory.
04:50I I I got it basically to, like, force myself to, like, use the lowest spec option. Are there any tweaks that you do still stand by that are the customizations that actually do matter? A couple of things.
04:59Like, we put a lot of time into our skills files and our, like, CloudMD. If I, like, open it up, like, you can you can see, like, this is probably a few 100 lines.
05:10There's, like, some interesting things in here. Like, it's say engineering practices. We're a startup.
05:15You're probably used to writing enterprise code, but that's not how we do things around here. And we have, like, a lot of things like that that we've, like, put into our CloudMD and our skills files over time. What else do I do?
05:25I always use fast mode. That's not a default. If you're trying to token max, like, you have to be in fast mode.
05:30I do use a context seven MCP. I think that's pretty helpful to get documentation. But other than that, like, use most of the most of the things out of the box.
05:39One core thing is that we always run Cloud and dangerously accept all permissions. Like, that is not the default, and that is the default way to run Cloud and Conductor.
05:48I think something that's really important to us is having, like, clear boundaries between well, we call them slot free zones.
05:56And having, like, parts of the code base or, like, parts of the documentation that we, like, know is written by a human. It's possible, like, that the a the AI can contribute to the the slot free zones, but, like, it has to be like, every line has to be read by a human. I think this actually like served us pretty well because if if you're not careful, the AI can like get in a a vicious a vicious cycle where like it sees bad code and then it writes more bad code as a result.
06:22And the same thing can happen in like the positive positive direction. We have like some lines in our code base that are like do not touch if you are an AI.
06:31Like this is for human eyes only. What's the conductor tech stack? It's a Tauri app.
06:35So it's using the native Safari web renderer. And the back end is technically Rust, but we write almost everything in TypeScript.
06:43So it's like probably 95% TypeScript on the, like, the desktop app. The web app is Elixir.
06:50It's a it's a Phoenix app. It's a very small app because literally all you can do in it right now is just like log in. But I am I'm a huge Elixir fan and I am always like pushing for more Elixir in our code base when we can.
07:02But most of what we're doing is in TypeScript. Another thing we talk about is like don't let the AI be your architect.
07:09Even the the concept of like a workspace here in the sidebar, which in some ways is just like an abstraction around a work tree, at least for right now. Like that's actually going to change soon.
07:20But even that, like, concept of a workspace, like, we as a human had to, like, think that through. The other thing is, like, like, design and, like, interface decisions.
07:30This concept of having, like, all your chats here on the left and then the chat in the middle and then the right sidebar where you can, like, review code changes or run your app, like, we put a lot of thought into into those decisions. And I think if you let the AI make your, like, UI choices for you, you can end up with something that, like it it just doesn't feel, like, crafted.
07:49And it's, like, really important to us that it it feels crafted. Like, even this decision. So, like, we we thought for a long time about how this open end button should work, which is kind of funny because now there's, like, so many apps have, like, have the same pattern.
08:01The thing that we were really thinking about is whether we should show the icons in the top. I was pretty against showing icons here at first because it just feels like, okay, in the top bar of our app, like, we're, like, advertising a different app.
08:16But now I I really like it and it's like a clear visual of, like, what's gonna happen when you click it. I think something we would do a bit differently is building the core of the app around, like, human written APIs and, like, contracts that the AI wouldn't contribute to as much.
08:30And then I think that, like, it's important to have a big chunks of your code base be like have, like, free rein for the AI where you can just, throw a ton of different ideas at it and know that it's, like, not gonna affect the core infrastructure. And I think right now the the boundaries are a little murky and like that's the thing we're we're working on improving.
08:48I think it's really important to us that we stay like a little ahead of the frontier. Like push push people's comfort zones a little bit more than they'd expect.
08:59When we first launched Conductor, most of the feedback we got was like, this is crazy. Like I barely can manage like one cloud code or like one codecs.
09:07Like how am I gonna manage like three or like even five? We also purposely made it so like you can't edit files directly. Like we like made it so that like anytime a workspace like has to be a work tree and it has to then create a PR and then you have to merge it.
09:22So we really like enforced our workflow. I think like what's what's exciting but also hard about where we're at is we we have to constantly adapt to where like the models are going.
09:32So that's one reason like we are putting so much work into like cloud right now is right now like you shut your laptop and the agents are gonna stop running. But like feels like we're very quickly moving to a world where the agents are gonna run for 10 times longer and they're gonna be 10 times smarter and they're gonna need to run-in an environment that, like, isn't constrained by, like, your Mac's, like, CPU.
09:52It seems like you're building Conductor in a very opinionated way. How do you build a conviction behind your decisions? K.
09:57That that's a great that's a great question because, yeah, like, it especially for, like, our audience, they want a lot of, like, configuration. And I do think it is important for the tool to, like, be flexible and to, like, feel like yours. But the way we build conviction is we force ourselves to use it because actually, we don't even force it.
10:14Like, we we just use it every day. And so if it doesn't feel right, like, we, like, quickly can can decide. But we we we're not big on analytics or, like, looking at, like, our a b testing or, like, it's very much a, like, gut feel.
10:29This feels right. Like, when I click this, it feels right that it opens in the center. And that way, like, I don't need a separate composer.
10:36And I can type messages here and it all feels unified. You sound like you default to Clog code in a lot of places, but Conductor supports Codecs too. Yeah.
10:44When do you reach for Codecs? Oh, yeah. I've recently actually been using Codecs more.
10:47Codecs is like the workhorse. It will power through like a specific problem or like it's not afraid to do a ton of tool calls and like debug something with me for a long time.
10:58Cloud, I'll reach for when I want a little more like back and forth. I feel like Opus is just like a little more creative, like a little more of a partner.
11:06And so I I would say like when I'm building out a new a new feature, like I I probably would like instinctively reach for Opus. And then when I'm like, okay. Now we just wanna get stuff done, like I'll go to Codex.
11:17Why isn't just a terminal good enough? There's a reason we moved from terminal interfaces to like GUI interfaces in the eighties.
11:27Like I think humans are spatial visual creatures and like having a a command line interface just like feels like it's feels very like restrictive. And I think it maybe works for the AI brains better than the human brains.
11:39But I think just like I wanna know that okay my chats are over here and my like review panel's here. I can talk to the AI in the middle. I just think like yeah.
11:47Bottom line like humans are like visual visual creatures. I also think like like out of like a like zooming in a little bit like there's a lot that you can't do in a terminal like that you can do with a
12:00user interface. Let's talk about token maxing. Yeah.
12:03What's your high watermark on lines of code in a day or spend in a month? I think the highest spend was when we were starting out Conductor, like in July 2025.
12:14I spent $22,000 on tokens that month. Granted that was with, like, previous generation of models.
12:21And the lines of code was must have been like tens of thousands that month. I'm very big on spending like on token maxing, like using fast mode, like think extra hard, high effort all the time.
12:33But we're not being on lines of code. We we try and keep the lines of code minimal actually. There's a bunch of reasons for this but I think like you can quickly spiral your code base can spiral out of control if you're like not careful about the lines of code added.
12:46But I I I think about it very differently if I'm like starting up an app versus like working in a established code base like Conductor. What's different about your workflows today from say six months ago? On a lot of like hard PRs, I would open an IDE and make changes by hand.
13:04And I also use GitHub, like the web app a lot a lot less now because I can just, like, review the code changes here in Conductor and, like, add comments here if I need to. We do have, like, a lot of PR checks that run. And so that's why we recently added this, like, this checks tab, which lets us just, like, add comments from GitHub, like, into Conductor.
13:24What's the most surprising thing you've seen someone else do with Conductor? One was, like, someone built, like, a mobile version of Conductor by, like, hacking together a bunch of our I don't actually even really know how it works, but I know it's, like, spoofing like IPC calls to our desktop app, which is pretty interesting.
13:41I think, honestly, Gary has shown us a lot of what you can do with Conductor.
13:47He is really putting it to the test. I think I've learned from him a bit about like how hard you can go on skills. Like skills are very much like a first class thing in in g stack and it's like it's there's some like interesting ideas there I think.
14:00Like especially around like onboarding. And we've added actually a specific mode for him called Gary mode, which by default does not collapse any of the tool calls. So you can see all the tool calls are default uncollapse.
14:13And you can even actually see Gary's face here if you're in Gary's mode. What feels obvious to you and your team that the rest of the world doesn't fully understand yet? Like I think there's like a lot of cool stuff to explore with like collaboration between humans and the AIs.
14:27Should you be able to communicate with sub agents? Should you be able to have like multiplayer chats where like multiple people are working on the same thing with the AIs? And then of course like the a metaphor we'll we'll often talk about is like feeling like the .
14:40You like wave the baton and like the instruments are playing in unison and then once in a while you wanna go to like the the trumpet player and be like, okay, you're out of tune. And then you wanna like zoom out to like the string section and like you should play a bit faster. But then most of the time you're like conducting at the the orchestra level.
14:58 in that like it used to be that code was the thing you were building. It was like the structure.
15:07You were putting time into, like, in in into, like, crafting the code. And now you're putting time into describing what you want and how you want it to be built.
15:16And the code is almost just like sawdust that comes out of that process. And like that leads to like a lot of like interesting conclusions. Like one of them is like, really what matters is your prompts.
15:26And like when the next generation of models come out, you can just like rerun your prompts again and then you'll get new code and the old code didn't really matter. I think that's one thing that like the world is slowly waking up to. I think like the submit a prompt, like the prompt request feature is sort of like an early experiment with malleable software.
15:44I the the metaphor that I always think of when I think of malleable software is like video games and how like when you play like Call of Duty, like the structure of the game is the same for everyone and like the skeleton is the same. But each person can like, I don't know, like use custom skins or like faster, like reload speeds or whatever.
16:02And like same way you can like mod a video game. I want you to be able to mod Conductor and like, yeah, build in your own workflows a little bit.
16:10It's important that like the structure feels the same and like people want software that's, like, been crafted and been, like, really thought through. But I also you know, like, video game mods make make the game feel more like your own, and I think that's gonna happen in the software as well.
The Hook
The bait, then the rug-pull.
Charlie Holtz sits at a desk with a split ergonomic keyboard, a gooseneck microphone, and a monitor running his own product. He co-founded Conductor — an app for orchestrating multiple Claude Code agents on a Mac — while in Y Combinator Summer 2024 batch. The setup tour that follows is really a philosophy lecture: by the end of 16 minutes, he has argued that code is becoming sawdust, and the thing worth building now is the system of prompts that generates it.
Named ideas worth stealing.
05:48concept
Slop-free zones
Human-only sections of a codebase. AI is barred from contributing without full human review. Breaks the feedback loop where poor code trains more poor code.
Steal forany project where AI writes more than half the code
15:00concept
The sawdust metaphor
Code is sawdust — a byproduct of the real craft. The prompt is the source of truth. Rerun on a new model, get new code.
Steal forreframing how you document and preserve an AI-assisted project
07:06concept
Dont let the AI be your architect
Architectural decisions, UI layout, and core API contracts must originate from humans. AI executes on top of those contracts.
Steal forany team that wants to maintain a crafted feel in an AI-heavy codebase
14:17concept
The conductor-of-an-orchestra metaphor
- Wave the baton at the orchestra level
- Zoom into one player needing correction
- Return to conducting at scale
Three zoom levels: directing all agents, directing one agent, merging output.
Steal fordesigning any multi-agent workflow UI or orchestration layer
15:48concept
Malleable software
Software users can mod like a video game — consistent crafted structure with personalizable workflows layered on top.
Steal forany product that wants to grow power-user behavior without compromising core UX
How they asked for the click.
VERBAL ASK
16:35subscribe
“WATCH NEXT / SUBSCRIBE (end card)”
Standard YC branded end card with full production credits. No verbal CTA.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
