07:04The Next New Thing · Interview
A free, open-source dictation app that beats the $15/month ones
Two hosts test FluidVoice against Wispr Flow and SuperWhisper live, side by side — and one of them deletes his paid app on camera.
July 4thThe argument in one line.
A wave of small open-source tools is patching specific weak points in AI coding agents — generic design output, computer control, parallel-agent management, subscription limits, and video comprehension — and builders who publish a personal fix well can turn it into a widely-forked project almost overnight.
Read if. Skip if.
READ IF YOU ARE…
- You build software with Claude, Codex, or another coding agent and are frustrated that its designs default to a generic, instantly-recognizable AI look.
- You run more than one coding agent or subscription and want a faster way to compare outputs or route between them without manual account-switching.
- You're job hunting and want to see how someone actually built and used an AI application workflow, including where it refused to inflate a fit.
- You want a fast, opinionated read on which of this week's trending GitHub repos are genuinely useful for AI-assisted building versus hype.
SKIP IF…
- You're not currently building anything with AI coding agents — this is developer-tooling news, not general AI news.
- You want a deep technical tutorial on any single repo — this is a reaction/roundup show, not a how-to.
The full version, fast.
This week's GitHub roundup groups tools around a common problem: general-purpose AI models are good but generic, so builders are shipping narrow fixes on top of them. Three design repos (Hallmark, Impeccable, Arcify) inject external style DNA, browser annotation, and diagram structure that Claude and Codex don't produce well alone. A geophysicist laid off in December built an AI job-search framework, now #1 trending, that scores job fit honestly rather than inflating every application, and it helped him get hired. The back half covers infrastructure for power users: giving agents real computer control, running several coding agents in parallel worktrees and merging the best result ("tokenmaxxing"), sandboxing AI-generated code for end users, and giving Claude the ability to watch and search video instead of just reading transcripts. The through-line: pick a handful of these narrow tools per workflow rather than expecting one model to do everything well.
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 – 00:26
01 · Cold open
Andrew previews the episode: three AI design repos, computer-control repos, a YouTube-learning agent, and a surprise guest. Sponsor tag for Zapier.
00:26 – 02:29
02 · Hallmark: design DNA cards
First design repo: feed it a website or image and it generates a "DNA card" documenting how to recreate that visual style, or an audit of an existing design. Adam likes having several of these installed to get different "designer perspectives."
02:29 – 04:04
03 · Impeccable: annotation + browser feedback
A design-language skill with a browser extension for circling an element on a live page and leaving a comment that routes back to the agent — a way to give feedback without knowing technical layout terms.
04:04 – 06:04
04 · Arcify: English-to-diagram
Converts a plain-English system description into a clean architecture diagram. Adam shares a personal story of Claude and Fable both failing to lay out a diagram the way he wanted the night before.
06:04 – 11:40
05 · AI Job Search: interview with the creator
Mads, a geophysicist laid off in December, joins live to explain the #1-trending AI job-search framework he built on parental leave: it scores job fit honestly, sometimes recommends against applying, and writes resumes/cover letters grounded in the user's real profile. Forks have been adapted for other countries' job markets.
11:40 – 14:19
06 · Desktop Commander: computer-control MCP
An MCP that gives an agent direct control of files and the terminal on your machine. Adam pushes back that Claude and Codex can already do most of this natively, questioning what unique value the tool adds beyond convenience shortcuts.
14:19 – 17:24
07 · Google's C++ libraries, weather tangent, and the Zapier MCP ad
A non-AI, all-C++ Google repo breaks into the trending list; Adam, a former C++ developer, is glad to see non-AI infrastructure work still thriving. Andrew flags an Iceland flood-warning tangent, thanks the audience for helping land the Matt Van Horn interview, then reads the Zapier MCP sponsor segment.
17:24 – 19:36
08 · Office CLI: agent control of Word, Excel, PowerPoint
Lets an agent operate Office documents with visual awareness ("eyes") so it avoids overlapping text and boxes — a common failure when agents edit these formats blind. Framed as a time/token saver versus asking Claude to hack it manually.
19:36 – 24:11
09 · Orca: parallel agents and "tokenmaxxing"
Runs multiple coding agents (Codex, Claude, others) side by side in isolated worktrees, each with its own mobile emulator, so you can compare and merge results. Discussion of "tokenmaxxing" and the value of asking for several divergent design variants instead of one.
24:11 – 25:47
10 · Cube Sandbox: hardware-isolated code execution
Boots an isolated VM in under 60 milliseconds so untrusted, AI-generated code (e.g. from end users of your own product) can run safely without reaching your database or other users' data.
25:47 – 27:55
11 · Herder: managing multiple agents from one terminal
A returning repo (second week trending) that brings a multi-tab, multi-agent interface into the terminal, where the newest coding-agent features land first but are hardest to manage across parallel sessions.
27:55 – 30:00
12 · OmniRoute: one endpoint, automatic model fallback
Routes requests across 200+ AI providers so a coding session doesn't stop dead when one subscription hits its usage cap. Adam cautions against switching agents mid-project without noticing, since each model needs to be talked to differently.
30:00 – 31:28
13 · Codex Plugin for Claude Code
OpenAI ships a plugin that brings Codex directly into Claude Code, letting developers stay in one tool while calling out to the other model for tasks it's better suited to, with OpenAI's own skill definitions guiding when to use it.
31:28 – 34:10
14 · Claude Video: giving Claude the ability to watch video
A repo that lets an agent see and hear video, not just read a transcript, so past recordings become searchable by visual moment as well as spoken content. Adam gets visibly excited about mining old course footage this way.
34:10 – 37:04
15 · Final thoughts and sign-off
Mads returns on camera to react to the full list and settles on Claude Video as most useful to him. Andrew closes by inviting viewers to reach out about collaborating on repo deep-dives, thanks Zapier again, and runs the subscribe/like/comment card.
Lines worth screenshotting.
- Standalone design skills exist specifically to break AI's tendency toward generic, "GPT-five-generated" looking output by injecting external design DNA instead of relying on the model's default aesthetic.
- A browser extension that lets you circle an element and leave a comment routes that feedback straight to the coding agent, closing a loop that's hard to describe in words like "div" or "h1" if you didn't build the page yourself.
- Diagram-generation tools solve a specific, repeated failure: general coding models default to sprawling horizontal flowcharts and can't compress structure into a dense, readable layout even after several rounds of correction.
- A geophysicist laid off in December, while on parental leave, built an AI job-search agent that evaluates fit against a posting, sometimes recommends against applying, and writes a resume and cover letter grounded only in the user's real profile.
- The job-search tool's honesty is a deliberate design choice — it will score a candidate as a poor fit and say so, rather than optimizing every application to look ideal.
- The project's creator says forks mattered to him more than stars, because a fork means someone is actually running their own copy of the tool, not just bookmarking it.
- People have forked the original job-search repo to work with the Danish, American, Vietnamese, and Brazilian job markets by swapping in different regional job-portal APIs.
- A general-purpose coding agent like Claude or Codex can already read and write files anywhere on a machine and operate a terminal, which makes a bolt-on "computer control" MCP redundant with capability the model already has rather than a new one.
- A trending, all-C++ GitHub repo with zero AI involvement broke what one host called a three-year streak of AI-dominated trending lists — systems-programming infrastructure work hasn't stopped, it's just been overshadowed.
- Feeding a coding agent a screenshot of the actual rendered output, not just the underlying code, is what fixes AI-generated documents where text and boxes silently overlap — visual review catches errors code review alone misses.
- Running the same build through multiple agents in parallel worktrees and picking the best result is being called "tokenmaxxing" — deliberately spending a token/subscription budget across parallel branches instead of iterating serially on one.
- Asking a single agent for several genuinely divergent design variants, rather than one "best" attempt, works because it's less about the model's creativity and more about giving the requester something concrete to react to and combine.
- Isolated micro-VMs that boot in under 60 milliseconds let a platform safely let end users — not just developers — run AI-generated code, because the sandbox boundary is what prevents database access or resource abuse, not trust in the code itself.
- Rotating between multiple paid AI subscriptions (e.g. three separate Claude accounts) to dodge weekly usage caps is common enough that dedicated routing tools now exist to automate the account-switching.
- OpenAI shipping a Codex plugin inside Claude Code — a competitor's tool built to run inside a rival's product — signals developers wanted cross-model access more than either company wanted brand exclusivity.
- A tool that lets an agent see and hear video, not just read its transcript, can locate the exact clip where something happened on screen — a transcript alone never records a visual reaction or an on-screen demonstration.
- Transcript-based video tools work well for talking-head content but miss most of the value in screen-recorded tutorials, where the information is what's on screen rather than what's said out loud.
Fix specific AI weak points with narrow tools instead of one do-everything model.
WHAT TO LEARN
Every tool in this roundup exists because a general coding model is good but not great at one specific thing — design taste, computer control, parallel comparison, safe execution, or video comprehension.
02Hallmark: design DNA cards
- Feeding an AI coding agent an external "design DNA" reference — a card documenting an existing site's palette, type, and layout — breaks its tendency to default to generic, instantly-recognizable output.
- Installing several competing design tools and switching between them, rather than relying on one, gives you a range of different "designer perspectives" to pick from.
03Impeccable: annotation + browser feedback
- A browser extension that lets you circle an element and comment on it routes non-technical feedback straight to a coding agent, useful when you can't describe a layout problem in technical terms.
- This kind of full feedback loop matters most when you didn't build the page yourself and don't know its underlying structure.
04Arcify: English-to-diagram
- When an AI keeps producing a flowchart or diagram in the wrong shape, a dedicated diagram tool that ingests plain English can succeed where repeated manual correction of a general model fails.
- Different general models vary noticeably in how well they compress a described structure into a dense, readable diagram.
05AI Job Search: interview with the creator
- An honest AI job-search tool that will tell you you're a poor fit for a role, rather than optimizing every application to look ideal, is a deliberate and rare design choice worth noticing when evaluating any AI tool that claims to help you.
- Forks are a more reliable adoption signal than stars, because a fork means someone actually installed and is running their own copy rather than just bookmarking the idea.
- A tool built to solve one person's specific problem, released openly, can be adapted by others for entirely different markets — regional job boards in this case — without the original creator doing that work.
06Desktop Commander: computer-control MCP
- Before adding a "computer control" tool to your AI stack, check whether your existing agent can already do the task natively — a dedicated tool's real value may just be convenience shortcuts, not new capability.
- General coding agents already have broad file-system and terminal access by default once you grant it; the gap for most users is knowing to ask, not a missing capability.
07Google's C++ libraries, weather tangent, and the Zapier MCP ad
- Not every trending repo is AI-related — traditional systems-programming infrastructure work continues underneath the current AI wave, even when it rarely gets attention.
- An MCP hub that lets you selectively grant an agent access to specific tools (rather than all-or-nothing) is a practical way to stay comfortable handing agents real permissions.
08Office CLI: agent control of Word, Excel, PowerPoint
- Giving a coding agent a screenshot of the actual rendered output, not just the underlying code, catches visual errors — like overlapping text and boxes — that code review alone misses.
- Legacy document formats like Excel and PowerPoint are token-expensive for a general agent to edit correctly; a specialized tool can produce better results faster and cheaper.
09Orca: parallel agents and "tokenmaxxing"
- Running the same task through multiple agents in parallel and keeping the best result trades extra token/subscription cost for meaningfully better output on high-stakes builds.
- Asking an agent for several genuinely different design or copy variants at once, rather than accepting its first attempt, produces better results because it gives you something concrete to react to and combine.
- You don't need a dedicated multi-agent tool to start — the same divergent-variant technique works inside a single regular chat with one agent.
10Cube Sandbox: hardware-isolated code execution
- A hardware-isolated sandbox that boots in under 60 milliseconds is what makes it safe to let end users — not just developers — run AI-generated code, since the isolation boundary is what prevents data or resource abuse.
- If your product lets users' AI-generated code run at all, that code should never have direct access to your production database or other users' data by default.
11Herder: managing multiple agents from one terminal
- The terminal remains where the newest coding-agent features land first, even though it's the hardest environment to manage several parallel agent sessions in without a UI layer on top.
12OmniRoute: one endpoint, automatic model fallback
- If you rely on more than one AI subscription to avoid usage caps, be deliberate about which model you're talking to at any moment, since each model responds best to a different communication style.
- Automating account-switching removes a real point of friction, but only pairs well with staying aware of which model picked up the conversation.
13Codex Plugin for Claude Code
- Every project benefits from working with more than one model rather than treating a single favorite as universally best.
- A plugin that lets one AI tool call out to a competing model for tasks it's specifically better at removes the manual overhead of switching tools entirely.
14Claude Video: giving Claude the ability to watch video
- A tool that lets an AI agent see and hear video, not just read a transcript, can locate the exact moment something happened on screen — information a text transcript never captures.
- Transcript-based video tools work well for talking-head content but miss most of the value in screen-recorded tutorials, where the information is what's shown, not what's said.
- Old recorded course or training footage that's gone unused because it's too long to scan manually becomes newly minable once an agent can search it by visual moment, not just spoken word.
Terms worth knowing.
- Design DNA card
- A generated summary (Hallmark's term) that documents the visual patterns — palette, layout, type — of an existing site or design so an AI agent can recreate or deliberately contrast that style instead of defaulting to a generic look.
- Tokenmaxxing
- Deliberately spending a large share of your AI token or subscription budget by running the same task through multiple agents or model instances in parallel, then keeping the best result, instead of iterating serially on a single run.
- Worktree
- An isolated copy of a codebase (a git feature) that lets a separate coding agent build, test, and run a variant of a project without interfering with other parallel variants.
- MCP (Model Context Protocol)
- A standard that lets an AI agent connect to external tools and services (files, apps, APIs) through one shared interface, so a single integration can expose many capabilities to any compatible agent.
- Hardware-isolated sandbox
- A lightweight virtual machine that runs untrusted, AI-generated code with strict limits on memory, network, and system access, so it can't reach a production database or another user's data even if the code misbehaves.
- Divergent design/variants
- A prompting approach where you explicitly ask an AI agent to produce several genuinely different options at once, rather than accepting its single first attempt, so you can compare and combine the strongest parts of each.
Things they pointed at.
03:20productClaude Design
05:19toolRoom GPT
05:30productCodex 5.6
05:35channelDan Shipper
11:40toolDesktop Commander MCP
13:50toolcaffeinate (macOS utility)
14:19toolGoogle C++ libraries repo
16:00linkMatt Van Horn interview (prior episode)
17:24toolOffice CLI
19:36toolOrca
24:11toolCube Sandbox (TencentCloud/CubeSandbox)
25:47toolHerder
27:55toolOmniRoute
30:00toolCodex Plugin for Claude Code (openai/codex-plugin-cc)
33:20productWideFrame
33:40productFluid Voice
34:50productDescript
Lines you could clip.
06:04
“A laid off geophysicist turned his own job hunt into an AI framework that got him hired for the very first time ever.”
09:10
“It's not about making up the best or the perfect candidate. It's always grounded in your own profile.”
13:40
“Have you tried Codex computer use? Like, it could do all this cool, wild stuff. What is this extra tool for?”
22:30
“Build me four different design variants. Give me four divergent examples of what we could build here.”
31:03
“Every project should work with more than one.”
33:40
“You don't see excitement in a transcript.”
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.
metaphoranalogystory
00:00Wanna help your AI create better designs? I've got three amazing GitHub repos for you. Want your agent to control your computer, Microsoft Word, Excel?
00:07I got repos that will do that and give it that kind of superpower. Are you learning a lot on YouTube? I've got an agent that will scrape YouTube, learn for you, organize everything, everything, and make it accessible to you.
00:16It is amazing. All that and a surprise guest coming up. Of course, we've got links to everything below.
00:21Let's get started. Presented by Zapier, the AI automation company. Alright, Adam.
00:27We've got three repos for design skills and then we're gonna go into the top 10 repos of the week. And the number one here when it comes to design is something called Hallmark, and you and I have experienced this.
00:36You ask your Claude agent or Hermes or whatever it is. Go create something for me, and it does, and it looks just like every other design that's out there. And so what this is, it's a way of changing your designs in a more creative way and you can do things like you can take any website or any image, give it to it, and then you'll end up with one of these cards.
00:58Where is that? Um, here. This is a card.
01:01You end up with one of these DNA cards as they call it, which tells you how to recreate it. So they did the same thing with Calendly. Or you can give it a design, and it will give you essentially a report card, an audit of what you can do to change, or you can use this to create from scratch.
01:14What do you think of this one?
01:16I like Yeah. I mean, this is one of the most frustrating parts of of building is that most of these agents, uh, they just won't give you on its own a great, unique, interesting design.
01:27And you're like, yeah. It doesn't look great, and you can fiddle with it, but it's not really getting better. And I think having some of this, like, other I'll use their words for it too.
01:35This other DNA to, like, mix in Mhmm. You know, we we did a little deep dive into this skill.
01:40And when you look in, there's a bunch of templates that are kind of built out. There's a bunch of components. It can build on top of these templates.
01:47It can also switch into this more custom mode and and bring all the components together in a in a new and unique way. I think there is something important to be said about trying to show yourself in a different way than the standard Clawd.
02:01Like, oh, yep. That's a GPT five five generated website. It's like, I've seen them a thousand times.
02:06I'll see them again. Or, oh, that's lovable. So I I like having these.
02:10I I like to to use not just one of these even, Andrew. I would install four or five of these Mhmm. And kind of switch between them to get a bunch of different designers' perspectives, basically.
02:20I've added these two types of cards to our collection here to help us understand it more. And one of them is this card that shows us about the creator. This is Asan.
02:30He's created multiple repos here that have done well, including this one, which I really like, Room GPT. Upload a photo of your room. It generates your dream room design with AI.
02:38Alright. That's the first one. The second one when it comes to design is called Impeccable.
02:42Pretty similar. It's a design language that makes your AI harness better at design. The thing that I like about this the most is this right here, where you can annotate a site and say, I like this.
02:53I don't like that. Claude Design has added this, but it took them a while to do it. This is a really nice touch that helps you interact with it in a in a, I think, much more human way.
03:04Yeah. You know, when when you don't have all the words necessarily, if you if you haven't thought through of, I know, I built the divs and the h one and h two and all the, like, technical words of how you lay out these web pages, it can be hard to give the agent feedback.
03:20Mhmm. So having a browser extension where you can circle something and add a comment right in your browser and make sure that that gets back to the agent. And that's the skill that help you build it.
03:30There's something about that, like, full loop that makes tools like this a lot easier to use, especially if you're not as familiar with how it was built or if you just Videcoded it and you didn't build it yourself and you let the agent build it anyway. So I I I love the browser extension.
03:46I really wanna test out this browser extension. I'm thinking of doing a deep dive on this one. If anyone knows a creator of it or if you're using this a lot, I'd love to do a deep dive and just spend maybe twenty minutes just going over it.
03:56It really is incredible,
03:58um, and I wanna understand it better. Alright. Next, final one for design.
04:02It's called Archify. You know when you have something that you wanna explain and you realize that an image is going to really explain it better than any words? And the problem that you have is how do I create an image that does it?
04:14What we have often seen is that flowcharts like this are the way to do it, but they're hard to put together. The idea behind Archify is that it will put this together for you, turn your English into easier to understand, um, images and workflows.
04:30And
04:31I get it. I have some hesitation. I wanna know what you think about this.
04:34Dude, I was struggling with this yesterday. Like, last night, I was on the couch trying to build something out, and I wanted a diagram to describe it. And Claude kept building it with, like, 12 steps in a horizontal order, so I had to zoom way in to see it.
04:48It's like, no. Like, this should kind of condense. There's, you know, a good a good chart here kind of enters at the top and has a couple different ways to move.
04:55And I'd say, no. I wanted to film the space more, and then it'd make a vertical version. Like, no.
05:00Not horizontal. Not vertical. Somewhere and I it just could not understand the layout that I was going for, and I ended up giving up.
05:07So I'm actually gonna take this one today and go and try and rebuild that chart. And based on what we've read here and seen, I bet it'll make a much more dense, easier to understand, easier to follow chart with, like, basically no work because I've already done the work to think through the the organization.
05:22Now I just need to display it nicely. So I I think these little tools, I love having a a bunch of those just kind of saved as skills, saved as tools so I can call them when I need them.
05:32You know, when, uh, Codex 5.6 came out, we looked at what the differences between that and Fable were. And Dan Shipper,
05:39I remember putting this in our video, showed how Fable does a better job of creating these types of flowcharts than Codex does. And I feel like there is some benefit to the model, but and I thought maybe this was a solved issue. But based on your experience and his experience with Codex, I guess it is not.
05:55Yeah. I was using Fable for this last night. So, uh, and it was it was a struggle.
06:00Okay. Those are three, what I call, like, buzz worthy repos. They're actually not from the top 10.
06:05They're from 11 through 20, but I picked them out because I thought they were really important for us to look at. Now let's look at the top number one of the week. It is AI job search.
06:14A laid off g geophysicist turned his own job hunt into an AI framework framework that got him hired for the very first time ever. We actually have that person.
06:25Matt, this is you. You built it. Hello.
06:28That's me.
06:30Is this true? Is this really what happened? You got laid off.
06:32You said I'm gonna use AI to solve my problem? Yeah. So I
06:36was impacted by layoffs mid December last year. And at the time, I was actually on parental leave.
06:44So when the baby was sleeping, I was fiddling around with the cloth coat myself. And, well, combining the two things, I thought this was a great opportunity for me to to try to well, I was also speaking with a a career adviser at the time, and I thought, like, okay.
07:00This is actually something that would, uh, that will work if you if you combine the things and the agentic workflows. So yeah.
07:08And creating the resume is one thing. Writing the cover letter, I demonstrated that before.
07:14It's impressive. It's the it's the application afterwards that really takes it to a next level. What happens after the creation of the letter and the,
07:23uh, and the resume? Well well, the idea is that you you fill in your profile, you, uh, you pointed at some some job posting and it it evaluates firstly, it evaluates your fit, right, to to that given description, and it's it's quite honest.
07:40Sometimes I've also refused jobs based on that criteria and the the scoring. And after that, it then goes and and tries to if if you proceed with the application, it tries to highlight all the things that you have done basically creating the best version of yourself for that given job.
07:59And it looks at the cover letter, tries to to find the different things that you have done that could be a great fit for the role. It looks at the firm that you're applying to do some research on on different news and stuff. And, yeah, then it it sends it back to you after this draft of reviewer version where it really tries to find all the things that you have done.
08:20And I think the most important thing is that it's grounded in your profile. Right? It's not about making up the best or the perfect candidate.
08:29It's always grounded in in your own profile. So
08:33How is it, by the way? I made this tool to make it easy for me to highlight, and I I wanna use it. Yeah.
08:37Uh, how does this scraping and searching work? Yeah. Um, so I've tried to make it as universal as possible in in the main repository.
08:48Mhmm. But I I did like how can I say I did adapt the the Danish version of it because I saw that other people had used different CLI tools with API calls to to the Danish job market?
09:02And there's been a lot of people that then embrace this repository and made their own version of the American market or the Vietnamese market or Brazilian market or whatever. So it's it's just a matter of of trying to find the ways whether these open portals for for finding job applications.
09:21Yeah.
09:23By the way, I've made that tool that I just showed you. I'm thinking I should put it on GitHub and make it public. It is on GitHub.
09:29That's how I get access to it on all the devices. You are someone who's created a few different things. Let me close out with that.
09:34What has it been like to put this out there and get the response from the world? Is it worth it when you're seeing, like, what, 13 stars, six stars? Well,
09:42I think it's well, I I I've never anticipated that any of these projects, not even the AI job search workflow, would be exploding the way that they've done.
09:55I mean Yeah. Back in mid March or something, I wrote, oh my god. There is 49 people that start this project, and it's amazing.
10:02And it's not only about the stars. It's also about the folks because that's actually people using it. So already at that point, I I was yeah.
10:10I was very enthusiastic. So, I mean, the way that it has exploded now, and it's probably also a lot of bots that have start the project at least.
10:19But yeah. I I still like, the the amount of response that I get, I think, is a true signal.
10:25Uh, people thanking me for open sourcing this, uh, people that landed that job. So I'm just I'm very overwhelmed at the moment with Yeah.
10:33Do people tell you when they when they get a job through this? Yeah. Yeah.
10:36They do. That's cool. Yeah.
10:38Oh, that is so exciting. Yeah. Alright.
10:40Mads, thanks so much for coming on here. I'm so excited about this and excited that you and I got to meet. Yeah.
10:44Thank you. We might wanna do a deep dive into this. Maybe you and I can follow-up afterwards.
10:48Folks, let me know if, uh, if you wanna see more about this. Maybe we'll do a a demo together. And if you know another creator, please let them know.
10:55Matt, did people tell you that you were on here?
10:57Uh, yes. There were some people that, uh, and, actually, I didn't know that I was number one trending. I mean, I was I am employed now, and, yes, I use my own workflow.
11:06Yeah. So I was just at work.
11:09I was like, why are so many people connecting with me on LinkedIn? And then I only actually, I only realized when I was like, in the weekend, there were people like, oh, you've been on this YouTube podcast and oh, sorry.
11:21On on this YouTube video presented with the the GitHub repo. I was like, okay. And then I went back and saw that, okay.
11:27I was actually number one trending. So that's that's how it went down. It's yeah.
11:30Oh, that's great. Yeah. Alright, Matt.
11:32Thanks so much for coming on here. This is fantastic. Thank you.
11:35Alright. Cool. And we should start linking everyone's LinkedIn profile.
11:39Thank you. Yeah. Alright.
11:40Thank you. Next. Bye.
11:42Bye. We're gonna go on to the number two of the week. Not so much as Cinderella story here.
11:47This is Desktop Commander. The idea behind this is and I think everyone could relate to it.
11:53You ask Claude to do something. It says great here. I'll give you, actually, an example for me.
11:57I like to find tweets that help me understand what people think about each one of these repos. I tell Claude, go do it. Claude comes back and says, well, actually, wasn't able to do this, but you can do it yourself.
12:06Here's a search. I'll go, freaking a no. You go do it.
12:09The same thing for anything else, especially if it has to do with my computer. I want the screen to be dim. It tells me how to do no.
12:15You go do it. The idea behind this is this is the you go do it tool that tells Claude how to go and do things on your computer and through this MCP gives it the power to do it.
12:25Um, what do you think of this one? Andrew, can't Claude already do all those things if you just ask it to do those things? It can read PDFs.
12:33It can use the terminal. It could open and and save files. Like, what is this doing that Claude or Codex can't do?
12:40I don't know. Can it? So I haven't I thought that with Claude, I can get it to open the files within the directory that I give it access to.
12:48I haven't told it to go beyond that directory. It can read and write anything. You can ask it.
12:53Say, hey. In my downloads folder, you know, there's this. Go and check, and it'll find your downloads folder and start looking too.
12:58Okay. How about the thing where if I want to shut my lid, I still want the computer to keep operating, I can go and put the terminal command in? Can it do that?
13:05You know, I bet if you ask Claude to do that, it would it would have a way to do that. Maybe this is, like, a bunch of shortcuts because it might take Claude a minute to figure out how to do that. Mhmm.
13:15Like, on Macs, there's this thing called caffeinate, which is the idea that the computer is incapable of going to sleep, you know, while some job is running. And so maybe this is like a bunch of shortcuts into tools like that.
13:27But I also guarantee Claude could figure out the caffeinate thing, you know, if if it hasn't already.
13:34Alright.
13:35Yeah. I'm not I'm not quite sure. Yeah.
13:37We went through this a lot before. I was super excited about this one until I talked with you, Adam, before we got started. I said, oh, actually Well, so I the models are plenty exciting enough.
13:46That's that's what I'm seeing. It's like, no. Like, have you tried Codex computer use?
13:50Like, it could do all this cool, wild stuff. Like, what what is this extra tool for? You know, eight months ago, I I would have been, you know, blown away.
13:58But now I I I think the built in tools are doing this. Alright. Fair point.
14:02And it did just take off. It was on a live, uh, feed somewhere. Woah.
14:08There's a, uh, what's it called, a flood threat. You know I'm supposed to go to Iceland today. I don't know how we're gonna get out of town here if the weather keeps getting this terrible.
14:15Alright.
14:16Seaplane, Andrew. Seaplane. Yeah.
14:19Look at this. I'm gonna get all these flood warnings here. This one is c plus plus building blocks that Google runs in production open now for anyone to use.
14:30It doesn't directly relate to me. Right? I'm not using c plus plus.
14:33What do think? As as a, like, a true nerd and c plus plus developer of old, I this warms my heart because this has nothing to do with AI.
14:43And it blows me away that there's something in the top 10 on GitHub that has nothing to do with AI. I think universally for the last three years, it's all been AI, AI, AI. And this is like this is a pretty old programming language.
14:56There's lots of versions of this language out there. They're all slightly differently capable of, you know, different access, different memory controls. This is like a pretty bare bones deep language.
15:06It's not a web language. And this is a bunch of people saying, hey. We want it to work slightly differently in, again, I'm sure very pedantic and nerdy ways that would be very fun to read about and mean nothing for AI.
15:18So I just love that, like, this stuff isn't dead. Coding isn't dead. People are out there, you know, building and developing the the infrastructure that all this other stuff lives on top of.
15:29So thank you to those people.
15:31Alright. I'm glad to hear you say that. By the way, I've said this.
15:35We asked for Matt Van Horn on here. You all helped out and supported it. We got the interview up with him, and one of the things that got me really excited was how many people came on with the first comments and said, got him.
15:47This is great. We worked together. I really appreciate all of you coming together to help me get Matt Van Horn.
15:51Um, and now the next thing that I'm gonna put out a call for is somebody to test out some of the repos and do this with me. Essentially, I'd like to do is what Adam and I did for Fluid Voice. We talked about Fluid Voice on the GitHub show a few weeks back.
16:06We got excited about it. Adam installed it, talked about how good it was. He said, you know what?
16:10I wanna do at least a quick one minute video talking about how good this is. And I said, you know, so many people like it. Let's do a ten minute video or whatever.
16:18We did it. Adam, the founder, has reached out to me so many times. He has commented on YouTube comments for it.
16:24He commented on Twitter comments that people commented because I posted the same video on Twitter, and people are thanking me for introducing them to it. I wanna do the same thing for some of these other repos. What else can we can we get other people to try together with me or with you on video here and, uh, and then expose the world to them?
16:42That's the idea. Alright. If that's you, reach out to me.
16:45Hi. At the next new thing.ai. You do know in a few hours I'm leaving for Iceland, but I will still have somebody here to reply to messages, and I wanna work with more people here.
16:53Alright. Speaking of the people who make this available to me is Zapier. Actually, the people who paid for the whole year of advertising because they believed that this thing that was nothing was gonna be something.
17:01Are Zapier, and one of the things that I wanna you about Zapier that I use is their MCP. I love that I can go to Zapier dot com slash MCP. I could say here's a collection of tools that I wanna give my agent, and I wanna give them this power but not that power.
17:13There's a thing that doesn't have an MCP. There's a thing that does. There's a thing that has a lot of power.
17:17I get it all, over 8,000 of them, and I get to pick what I give access to, and then I connect the agent to it. And I do this anytime there's a new agent that I wanna play around with so that I can give it the power that it needs and restrict it so that I feel as safe as I always wanna be.
17:32Alright. Go to zapier.com/mcp, and please start telling them.
17:35Because you know what, Adam? They're telling me that everyone's saying that Matthew Berman sent them over. The other creator who was on here, I've been talking with him.
17:42He's a really nice guy, and I appreciate it. I'd love for people to say, hey, Zapier. Thanks for taking a shot on Andrew and the team.
17:48And, yeah, I checked you out because of them. Alright. Let's move on.
17:53Next is Office CLI. This lets your agent operate Word, Excel, and PowerPoint, but it's more than operate it. You and I both know that when you have any creative ability being put in the hands of AI, there's a little bit of sloppiness because you can't fully see it.
18:07And so it puts text on top of other text. It puts a box on top of something else, a chart on top of some of another chart. This has eyes, different.
18:14And that's what allows it to, um, that's what allows it to do a better job of controlling these tools. What do you think?
18:21Yeah. The the screenshot thing seems to be a trend lately. You really wanna give the models
18:27the view of what you're seeing, not just the code, not just the the bits and bobs. And and that does really seem to be effective. I'm I do this naturally anyway, take lots of screenshots and funnel them into Cloud Code and Codecs.
18:40Right. You know, what I the other piece, though, that I see here is is, man, these older tools, Word, Excel, PowerPoint, they're hard to interact with.
18:49And they've always been hard to interact with. It's so difficult to programmatically create between an agent and Excel like, Claude's going to figure it out, by the way.
19:01Claude will figure out how to edit your Excel file if you ask it to. But it's going to spend a lot of tokens to do that, which means a lot of time. And you're using up your token budget for it.
19:12And so having a tool like this that can make give you basically better results a lot faster, especially in PowerPoint. I've never seen it make a great PowerPoint for me.
19:22So, yeah, these are cool. Again, this fits in with the trend of, you know, you should just have five or six of these tools. If you're constantly in PowerPoint, you should have a couple PowerPoint tools that really supercharge your access to the application.
19:36Alright. Next. Orca.
19:40Run five coding agents side by side, then merge the one that got it right. This is something that I think is best looked at here to understand it.
19:50This is what it looks like, and this is some of what it does, some of the features. I love anything with a mobile companion.
19:57You introduced me to Versus Code. I think their mobile companion is pretty terrible. Um, I want I want something that has a more mobile access.
20:05Yeah. I love this design mode, which you and I just talked about. What do you think of this one?
20:10It's a cool it's a it's a cool concept. It fits in with token maxing, which is this trend of, can I spend a 100% of my token budget from Codex and from Claude and, you know, from wherever else you're buying it?
20:24And I've got three accounts that I'm spending on all of them. This fits because part of the premise here is we'll put Codex on it and put Claude on it. And maybe even for each of those, try two different versions and create four fully functioning copies of the website or app or whatever it is that you wanna use, and then you pick which one you like.
20:43That's pretty hard to manage if you're just in Versus Code or if you're just in the terminal. So it's really nice to have one place to drop all of these where you can maintain these different parallel conversations. Because this thing, it says that each work tree gets its own emulator.
20:59What that means is it's actually going to run four copies of your app next to each other. And it's going to code it four times. And, man, I don't know.
21:07And if you didn't have an interface for that, if you didn't have a tool like this, I don't know how you'd be able to think through all of that and manage it without the agents running into each other. Knowing what I'm building, things like that toolbar that you just saw me use Yeah. Is this something that I should be using?
21:22Yeah. It it's pretty good practice when you're building something like that. Mhmm.
21:27Even if you're just talking to one agent about it. Say, build me four different design variants.
21:32Give me four divergent examples of what we could build here.
21:37Because what happens is the agent's not all that creative, and you may or may not know what you want. So if you ask the agent to get creative to produce divergent examples, now you can pick.
21:47It's, oh, I love this part of option b. I love this part of option c. And that's really it's it's not about one of them being the best.
21:54It's about stimulating your creativity. You get to see which ones you like, and then you combine the best parts of each of these into one. So that's that's how I'm building anything.
22:04And, again, this tool is . But just to start, try it in regular Claude. Use the word divergent and ask for ask for four mock ups instead of one.
22:17Okay. Alright. And what about this?
22:21We have seen people say that Codex five six is better at writing. That Fable is better at creating a software with one shot.
22:30Is this what I might use if I wanted to work with both of them, for example, and assign codex the writing of a website and all the copy, ask Fable to build the software, etcetera? Yeah. This would work for that.
22:41Yes.
22:41And and that's exactly what we're doing too. You know, we're writing all our copy in in one of the GPT models. The the new one, we've we've been playing a lot with five six, and it's like, it's better at design, but still far worse than Fable and others.
22:54So, yeah, it's it's kind of important to have your favorites. But, Andrew, I I still love to pit them against each other. I love to have GPT look at the the Claude produced design and give feedback.
23:05And you just get different ideas. Right? You kinda ,
23:09and and you get to to mix it up a little bit. So you're not quite so formulaic. When you say we, you mean you all at Gateway X.
23:15You're creating that software. In addition to backing entrepreneurs, you're building software for yourselves or software to eventually sell or what?
23:23Yes.
23:24Both of those things. We build a lot of software for ourselves. We build meeting tools.
23:28We build transcript analyzers. We build sourcing tools. We're looking constantly for, like, people that could be great founders of businesses like this.
23:37And, you know, so we build all these tools. We're not just sitting here waiting for people to show up.
23:42I see. Everything we do is is is AI native. Everything is AI first.
23:47Meaning, you might build a software yourself internally and then find a founder who's better at taking it to market and Yeah. A lot of the the times, that that's how it works. You know?
23:55Instead of instead of a a business plan or something, no. Here's a fully functional prototype. Come and look at this fully functional prototype, and the right person looks at it and says, oh, I know who I'd sell that to.
24:07Or, oh, I'd love that, but for doctors instead.
24:11Yeah. Okay. Let's go on to the next one.
24:14Number six for the week is KubeSandbox. It boots a hardware isolated VM in under sixty milliseconds so agents can run untrusted code safely. I've asked you to explain it here to me so that I understand this one.
24:26I don't know it. This is another good kind of nerdy technical one.
24:31AI can write code, and a lot of us are writing code with with AI. But the users of our applications are typically not writing code.
24:42If you build an analytics platform, you're probably going to put up, here's the charts that people can see. There's an AI agent.
24:48But can that AI agent that you're letting your users access, can it write Python? Can it scrape the internet? Can it analyze whatever random data they send in?
24:58No. You don't let them do that because it's not safe to let someone run their own code, especially AI generated code, on your server. How can you be sure that it won't access the database or steal other users' data or use all of your bandwidth?
25:12So these sandboxes are the way to make that work. You'd say, well, I'm not going to give you access to the full system.
25:19Your agent can write code, but it can only run code within these very particular walls. I'm only going give you a little bit of RAM.
25:28Maybe I won't give you access to the internet. Definitely, there's no access to the database or anything else. But if you upload your files and you want to analyze them, sure.
25:36Write Python and and analyze away.
25:39Okay. I wonder, by the way, as I've been flipping through this, is this helpful for people that I'm showing all this different all these different, uh, pages with context? Let me know in the comments.
25:48And, of course, those of you reaching out to me directly, let me know that way too. Alright. Herder is back.
25:53It lets you run a herd of coding agents from one terminal. We've talked about this before. This is the second time in a row, I think, that it's hit the top 10 list of the week.
26:02I like this tweet. Last time I showed you this video of the setup, and I thought it was very involved in technical. I like how fast jills, uh, said it in a tweet.
26:11I'm actually gonna zoom in and get him to one and a half minutes into this here.
26:26I can't hear the audio, Andrew. I'd be able to You cannot hear the audio?
26:30Okay. Alright. I'll just, uh, I'll leave it here in the chat in the document, which, of course, is linked to below for people to see it.
26:37The the idea here is that it's built to enable you to work with multiple agents, know when they're done, know what project you've given them. What what do you think about this?
26:47Yeah. I mean, we're I love it is the short answer. Uh, we get a lot of this experience if you're using Versus Code and the and the plug ins there, the extensions that let you see these in multiple tabs, run multiple agents at one time.
27:02It's really hard to do that in the terminal. And honestly, the terminal is a much more powerful way to use these coding agents. They're the features that come out, they they're there first.
27:12And if you go and look, probably half the features almost never make it to the extension anyway. So if you really want to be a deep power user, you're in the terminal working with this stuff. Oh, then it's really hard to run multiple at the same time.
27:24It's hard to do what we've been talking about where you're saying, oh, I want to see five versions. I want them to work in parallel. And I want to research this at the same time.
27:32Wow. That's, like, hard to manage. Now you've got a bunch of terminal windows.
27:36You can't keep track of them because they're all just white text on a black background. So it's tools like this that bring some of that user interface from Versus code or cursor or wherever else and bring them into the terminal so that you get some of those same kind of comforts and the ability to switch back and forth.
27:55Next, OmniRoot. We've talked about this one too. One endpoint, 250 AI providers automatic fallback.
28:01Here, what you're doing is you're coding with an agent, and we all have seen the experience where it'll say you've used 80% of your allotted usage for the five hour window. You've used 90% for the week, etcetera. And then boom, you're just done.
28:15And now you have to go and figure out what do I do next? How do I switch? And even that little bit of pain is a distraction.
28:20What this will do is it will automatically switch you off from one to the next so that you can continue working. And these tools have become more and more popular.
28:30Um, do we have anything new to add on to this since last time? Maybe just a recommendation,
28:36which is we've already talked about some of these coding agents are better or worse at certain tasks, whether it's design or debugging or computer use. Mhmm.
28:46And so you can't really just switch between them willy nilly. Using a tool like this is very valuable if you say, I've got three Claude subscriptions that I need to be able to rotate through. But I wouldn't, in the middle of a project, want to randomly switch between Claude and Codex without realizing it because I talk to them differently.
29:05They talk to me differently. And so this is actually a way where you'd say, well, wait. I've got three of these subscriptions.
29:11I got three of these. And it brings all that control into one spot. But still be aware of which model you're using at any given point so that you know how to talk to it and you know what kind of task to give it.
29:23Alright. I wonder, like, how many people are using this type of setup to switch between multiple subscriptions.
29:31Yeah. Honestly, I'm still doing it manually. I just when I run out, I log out and log in to another one.
29:38You know what I'm seeing is someone will have, like, a spouse or an employee who really is using ChatGPT just to chat, and they have the subscription. They can do things like code with it. They can do things like use an agent with it, but they're not.
29:52So they just give the API access or they give they give access to it, and the person who's one of these token maxers gets to use it. Makes sense. Alright.
30:01Next, Codex plug in CC. This is something that basically OpenAI has recognized that people are using Cloud Code, of course, to develop.
30:12They don't wanna switch away, but they might want some of the OpenAI models more within reach. And so they're basically building codecs into, uh, into Cloud Code.
30:23What do you think? Yeah. This is awesome.
30:25Uh, if if you're not working in this way, you should. Download download these skills here. You should be cross pollinating ideas from both of these and getting the best out of both models.
30:37So you wouldn't say, I I build with Fable because Fable's the best. No. Fable's really good at certain things.
30:43In GPT 5.6, it's gonna be even better at other things. And so every project should work with more than one. But it's hard.
30:51I mean, this is a theme this week. It's hard to manage all these agents working together, sometimes working in parallel, sometimes working on the same file.
31:00And having the skill means that now you you can still just talk to Claude Code, and it will call out to Codex for the tasks that you've asked it to call out for.
31:12And OpenAI has done a good job in the skill definitions here to tell it when that's a good idea. Here's the things we're good at. Here's how you should use me.
31:21And so now you get to use both models, but you don't have to figure out these parallel tools that we've been talking about.
31:28All right. Final one. Claude video, we looked at this last week.
31:33Again, it's coming in. This is a phenomenal project. What this does is it allows Claude or whatever agent you're using more and more, by the way, we're using Claude to just to define the whole category like a Kleenex, and I know that it's just not right.
31:50But I I also feel weird saying things like your agent or Claude and Codex and all these other models. Anyway, what it does is it basically allows them to see and listen to video. And as a result, you can both edit video through this type of thing, but more importantly, you can learn from video.
32:06I've seen people use it to learn because they're doing research projects. People use it to save it just because they want everything that they watch in videos to be searchable so that they could say, you know what? I know I saw someone do a thing about this.
32:17Where was it? Well, you come back into your agent and you say, go find me where where I I studied that and you get the part of the video and you get the thing that you learned.
32:24I I see that happen to me all the time. So we've got the videos for it here. Anything new to add on this one?
32:32We we've got a course that we used to teach. Mhmm. It was Bootstrap Giants.
32:37There was a sales accelerator. A bunch of people went through this, and it was a great course. It's all recorded.
32:42We've tried to repurpose the content a lot of times, and we've struggled because we've got the resources. We've got the transcripts, and we've got the videos.
32:51But the videos are three hours long. Like, it's hard to work through. And so, honestly, we've never done it.
32:57I I'm kind of feeling inspired right now to use this tool to go back and see if, oh, some of the most interesting visual moments, what was the text happening at that time? And also, some of the most interesting moments in the transcript, what was happening in the video, I think we might finally be able to mine those for some interesting ways to recut that same content.
33:19So I'm I've I've got a little inspiration for that. I wanna do it. It's honestly it would just not it would just not be possible without something like this.
33:27Yeah. I'm finding that a lot of video editors now that work with agents are basically looking at the transcripts like Descript does, and it's good but only for talking head videos. Yeah.
33:37A lot of what you're doing really is on the screen. You know? That's right.
33:41Like, I I don't just wanna know how a thing works. I wanna know the moment where they showed it work on the screen, and that's the part that I wanna watch for example. Like, oh, that was his face.
33:49He showed excitement there. Right. You don't see excitement in a transcript.
33:54You know, the transcript of this moment isn't gonna say, Adam was excited, and he put his arms up. Yeah. Right.
33:59Exactly. Oh, it's gonna be so obvious on my face and with my body. And and so, yeah, I I think there's there's, like, real potential in this.
34:08I'm I'm excited to try it.
34:09Alright. That's the top 10 mads. Are you still watching us live?
34:13Can I bring you back on camera? Yes. He said yes.
34:16Let's bring him back on. Can you still hear me? Yeah.
34:20I thought you were gonna drop out. I'm excited that you watched the whole thing. So now that you've seen the whole thing, what do you think of this?
34:25Especially with the last one, I think I'm gonna go with that one.
34:28That's for sure. There you go. What do even use it for?
34:33I don't know yet, but I I will I will definitely use it, I think, because well, I could just relate to it in in other personal projects where I've used these YouTube transcription Python workflows and stuff like that to, for example, build up knowledge databases and stuff like that, then it would be great also to to have the video established for that.
34:57So yeah.
34:59You know what? I've used an app called WideFrame that Gateway X introduced me to Mhmm. Adam's company.
35:05And it is amazing. I will give it all the video that I shot on my action camera, and I'll say, what's the story I could tell with this?
35:13And it's amazing where it says there was a lot of laughter in this moment, or there was excitement and sliding in that moment. I think this would work. And then I said, but I need drama.
35:21It goes, okay. I found some clips where people fell. And so that kind of interaction is really helpful.
35:26Now this only works for something that's on my desktop, which is what I like about wide frame. Mhmm. But it will not scrape YouTube.
35:32It will not go to Loom videos. I forget what we were using, Adam. And wide frame is not cheap, Andrew.
35:38I I don't remember how much it it costs, but it is it's a it's a real it's a professional tool. It is professional tool for professionals.
35:45And so that's I mean, this is actually a great way to kinda sell this one then too. Hey. You could do the thing that the professionals are doing, but just in your regular Clod subscription.
35:54You don't you don't have to pay for anything extra.
35:56That's awesome.
35:58Alright. Mads reached out to me. I hope all of you will reach out to me.
36:01I don't care nearly as much about subscriptions, likes, and comments, so I do feel good about this. Everyone now is kinda making fun of me in in chat about it. Um, I definitely do feel good about it.
36:10I definitely will be checking it instead of, I don't know what, um, paying attention to what's going on in Iceland. Having said that, connecting with Mads, connecting with you all is way more exciting for me. So if you're out there, find a way to connect with me.
36:21You got my email address, hi@thenextnewthing.ai, or, um, and it it will be in the description. Find a way to connect with me.
36:28If you see us talking about something, let the creator know just like Matt. He didn't even know he's number one, and he was number one. Let them know let them know that they were on here, and, uh, and I I'm really excited about what we're all building here together.
36:40It's like this guy, he's created something for himself that ended up touching thousands of people. All this software is going to change people, and all of us building it is exciting. Look at me.
36:49Look at how excited I am with this with this little thing, and it's super freaking helpful. I love your little extension. I wanna try it too.
36:55Alright. See you all in the chats. And, uh, actually, I've got another collection here that you really need to see.
37:01Click the link, and I'll see you in there. Bye.
The Hook
The bait, then the rug-pull.
Andrew opens by promising three GitHub repos that fix AI's generic design instincts, tools that hand a coding agent real control of your computer and Office apps, and a repo that scrapes and organizes everything you learn on YouTube — then teases a surprise guest before getting into the list.
Named ideas worth stealing.
01:12concept
Design DNA extraction (Hallmark)
- Point it at an existing site or image
- Generate a DNA card documenting the style
- Apply that DNA to a new build, or audit an existing one against it
A repeatable way to give an AI agent an explicit external style reference instead of letting it default to generic AI-generated design.
Steal forany project brief you hand an AI designer, so it stops producing the instantly-recognizable default look
22:30concept
Divergent design variants
- Explicitly ask for 3-4+ options, not one
- Use the word "divergent" in the prompt
- Combine the strongest parts of each into a final version
A prompting habit for countering an agent's tendency to settle for one "good enough" output by forcing genuinely different attempts up front.
Steal forany AI design, copy, or layout brainstorm where you'd otherwise accept the first result
19:55concept
Tokenmaxxing
Running the same task through multiple agents or model instances in parallel — sometimes several accounts of the same model plus a competing model — and keeping the best output rather than iterating serially on one.
Steal forhigh-stakes builds where output quality matters more than the extra token/subscription spend
31:03concept
Model specialization split
- Assign copywriting to one model
- Assign the software build to a different model
- Cross-check one model's output using the other for feedback
Deliberately using different models for different parts of the same project based on where each one is comparatively stronger, instead of picking a single favorite.
Steal forteams building both marketing copy and product simultaneously
How they asked for the click.
VERBAL ASK
17:10product
“Go to zapier.com/mcp, and please start telling them... I'd love for people to say, hey, Zapier, thanks for taking a shot on Andrew and the team.”
Mid-episode sponsor read woven into the show's own topic (MCP tooling), paired with a personal usage anecdote and an explicit ask for the audience to publicly credit the sponsor by name — softer than a pre-roll ad because it reuses the episode's own subject matter.
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.
