Building a Real AI Jarvis with Cursor and GPT Realtime 2
A step-by-step build of a talking, screen-controlling desktop AI companion — no coding experience required, just long prompts into Cursor.
Posted
5 days ago
Duration
Format
Tutorial
hype
Views
21.7K
636 likes
Big Idea
The argument in one line.
A complete voice-controlled AI desktop companion — with animated face, web search, screen control, and image generation — can be built by a complete coding novice through a handful of long, detailed natural-language prompts into an AI coding tool.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You want a working example of how far a single well-written prompt can carry an AI coding tool before any manual code editing is needed.
You're curious what a voice-first agent (as opposed to a chat-first one) actually looks and feels like in daily use.
You want to see computer-use / screen-control agents demonstrated on a real desktop, including their current rough edges.
You're evaluating GPT Realtime 2 specifically and want to see its latency, interruption handling, and tool-calling in an unscripted demo.
SKIP IF…
You're looking for a rigorous engineering breakdown of the underlying code — this is a live prompting demo, not a code walkthrough.
You want a finished, polished product — the agent visibly breaks (parse errors, missing API keys, confirmation loops) multiple times on camera.
TL;DR
The full version, fast.
The video demonstrates building a voice-based desktop AI companion end-to-end using only natural-language prompts into Cursor, no manual coding. The core mechanism is iterative: one long initial prompt describing the vision (voice conversation via GPT Realtime 2, animated face, artifact panel, tool use) produces a working app in about ten to fifteen minutes, and each subsequent round of feedback is delivered as another long prompt bundling multiple fixes and features at once — API key wiring, visual design changes, computer-use mode, and an image-generation and editing workflow. The agent can search the web, render diagrams, generate and iteratively edit images in a numbered grid, and take control of the desktop to open and type into other applications. The main actionable conclusion is that the bottleneck in this workflow is prompt specificity, not coding skill: vague requests produce broken behavior (crashed diagrams, silent tool failures from a missing API key, repeated confirmation prompts blocking screen actions), while long, itemized prompts covering UI, behavior, and edge cases in one shot produce large usable jumps in functionality.
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.
Live voice exchange with the finished 'Jarvi' companion — menu, mood changes, computer control — shown before any explanation of how it was built.
00:39 – 01:46
02 · Introduction & premise
Riley explains this is a 3-prompt build in Cursor using GPT Realtime 2, and that zero coding experience is required.
01:46 – 02:48
03 · Step 1: Download Cursor & new project
Downloading Cursor, opening the agents window, and creating a new 'Riley Custom Jarvis' project folder.
02:48 – 05:49
04 · Step 2: First prompt & OpenAI API key
Riley drafts the master prompt describing the desktop AI companion (voice, animated face, tools) and creates an OpenAI API key to paste in before running it.
05:49 – 08:01
05 · Step 3: Meet Ricky
Cursor finishes generating the app (~10-15 min); Ricky's animated face and basic voice capabilities are demoed, including a mermaid diagram that initially fails to parse.
08:01 – 09:08
06 · Sponsor break: Agents for Business
Riley pitches his team's service helping companies deploy shared AI agents across Slack and other tools.
09:08 – 09:48
07 · Step 4: Second prompt (10 changes)
A single bundled prompt fixes the mermaid parser, adds the Exa API key for web search, and redesigns the UI to a simpler 50/50 face-and-artifact layout.
09:48 – 11:26
08 · Step 5: Exa key & re-test
Exa API key created and pasted in; the redesigned Ricky is tested with mermaid charts, editing an existing chart, and a web search about the World Cup.
11:26 – 12:16
09 · Step 6: Third prompt — search rendering & menu
Prompt requests cleaner streaming markdown for search results, a spoken 'show me the menu' command, and better mouth-animation sync.
12:16 – 14:36
10 · Step 7: Computer use mode
Ricky is switched into computer-use mode to open and type into other apps (Codex); per-action confirmation prompts are identified as friction and the builder asks to reduce them, plus a request for a minimized, translucent computer-use window.
14:36 – 18:43
11 · Step 8: AI image & thumbnail generator
New capability: generate and iteratively edit thumbnail images of Riley in a numbered 3-wide grid, referencing a photo of himself, with voice commands like 'edit number 2' and 'pull up photo 15'.
18:43 – 21:08
12 · Final demo & close
A last mermaid diagram and web search test, then computer-use mode again; Riley wraps with the GitHub repo, the full prompt in the description, and a call for viewers to build their own versions.
Atomic Insights
Lines worth screenshotting.
A functional voice AI companion with an animated face, tool use, and screen control was built through natural-language prompting alone, with zero manual code written on camera.
The first prompt to Cursor took Cursor roughly ten to fifteen minutes to turn into a working desktop app.
Bundling ten distinct change requests into a single prompt (bug fixes, new API keys, and a full visual redesign) was handled by the coding agent in one pass.
A missing third-party API key (Exa) caused the voice agent to fail web search silently rather than surfacing a clear error to the user.
The computer-use mode required explicit per-action confirmation before typing or pressing Enter, which the builder found disruptive enough to specifically prompt the agent to remove.
Image generation was designed as a numbered grid (new images always land in position one) so that a spoken command like 'edit number two' has an unambiguous target without any pointing or clicking.
A single sentence describing a UI layout change (50/50 split between an animated face and an artifact panel, simplified buttons) was enough for the coding agent to execute the redesign.
The same confirmation-based safety gate that stopped the agent from typing without permission also blocked it from submitting text with Enter, requiring a second explicit approval for what the user considered one action.
Voice interruption (talking over the agent mid-response) worked well enough in the demo that the builder treated it as expected baseline behavior rather than calling it out as a feature.
Takeaway
Prompt specificity, not coding skill, is the real bottleneck
WHAT TO LEARN
Building a working voice AI agent required zero manual code, but every leap in capability came from one long, itemized prompt rather than many small ones.
A single detailed prompt describing the full vision — behavior, visual style, and tool access — can produce a working application in around ten to fifteen minutes without any manual coding.
Batching many distinct changes (bug fixes, new integrations, and a full visual redesign) into one prompt is more efficient than fixing issues one at a time, and modern coding agents can execute all of them in a single pass.
Missing a required API key doesn't always produce a clear error — a voice agent can fail a task silently and just report vague uncertainty instead of stating the actual cause.
Safety confirmations on agent actions (like typing or pressing Enter) are valuable but can compound into repeated friction; designing when confirmation is required is itself a decision worth making deliberately, not defaulting to maximum caution everywhere.
Giving an AI system a small, unambiguous addressing scheme (like a numbered grid where the newest item is always in a fixed position) removes the need for pointing, clicking, or vague references when working by voice alone.
Glossary
Terms worth knowing.
GPT Realtime 2
OpenAI's voice model used for live, low-latency spoken conversation, including the ability to be interrupted mid-response and to call tools while talking.
Cursor
An AI-assisted code editor that can generate and modify an entire application from natural-language prompts, used here as the sole tool to build the desktop app.
Computer use mode
An agent capability that lets the AI directly control the desktop — opening applications, typing text, and pressing keys — instead of only responding in a chat or voice interface.
Artifact panel
A side panel in the app's UI where the AI displays generated content such as web search results, charts, and images, separate from the conversational voice interface.
Exa
A web search API built for AI agents, used here to give the voice companion the ability to search the internet and summarize results.
Mermaid diagram
A text-based diagramming syntax that renders as a flowchart or chart; used by the agent to visually explain concepts like how AI systems work.
“The answer is zero out of ten. You can be a complete novice to coding, and you'll be able to create a voice agent for your personal use or for your business.”
direct, confident claim about accessibility→ TikTok hook↗ Tweet quote
15:08
“So I could say edit number two and make it blah blah blah, and then that one, the newest one will always show up in position one.”
concrete explanation of a reusable UX pattern→ IG reel cold open↗ Tweet quote
13:19
“Is there a way to turn off that explicit approval mode? Are you able to do that? If so, do it.”
shows the real friction point of computer-use agents in one line→ newsletter pull-quote↗ Tweet quote
The Script
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.
17px
analogystory
00:00Hello there, Jarvi. How's it going? Hey there.
00:03It's going smoothly on my side. Jarvi, can you please open the menu? Sure.
00:08Let me flip the menu now. Change the background to pink?
00:36I want you to open up the browser. Okay. I'll open your default browser so we can use it.
00:42The browser is open and ready for your next step. Hey. Can you please open up Codecs?
00:47Okay. Let's bring up Codecs for you now. Alright.
00:50So this right here is my mini Jarvis that I created in three prompts using Cursor. And this AI is connected to GPT Real Time two, which is OpenAI's new voice model. In this video, we're gonna be building a custom Jarvis, and this is gonna be a full voice agent that can use your tools, and it will be able to fully control your computer.
01:13And for this, the tool that we're gonna be using is Cursor. And you might be wondering how much coding experience is required.
01:20Well, the answer is zero out of 10. You can be a complete novice to coding, and you'll be able to create a voice agent for your personal use or for your business.
01:31But the really magical technology behind Jarvis is OpenAI's brand new GPT real time two, and it is so realistic to talk to.
01:42The AI agent will just speak to you in real time, and we're about to create that right now. So in order to get started, you should download Cursor if you haven't already, and you can find it at cursor.com. And you can very easily download it for Mac OS, and then you're just going to sign in.
01:59When you first open up Cursor, it may not look like this. At the top right, it'll say open agents window. That's I'm gonna be using.
02:06It's just a different view of the same technology, and we're gonna be using the agents window. So it should look something like this.
02:13And all we're gonna do to start out is create a new project. I am going to come to this workspaces tab.
02:20And here you won't see anything if it's your first time using Cursor. And I'm just going to hover over this and click open workspace. And I'm gonna hit open folder.
02:29I'm gonna go into my desktop for now. And we're just going to create Riley Custom Jarvis.
02:36And I'm going to create this folder, Riley Custom Jarvis, and press open. Now I am creating an app at this folder location on my computer, Riley Custom Jarvis.
02:49Okay. So now it's time to plan our first prompt. We are going to give a really long prompt to Cursor, and it's just going to build Jarvis for us.
02:57And we can use any AI model model to create it. What's important for us here is we need to give it an OpenAI API key. And in order to find your OpenAI API key, you need to go to openai.com/api-keys.
03:11And so we are going to create an API key. And remember, keep this absolutely safe, and I'm just gonna create a new secret key now. Create new secret key.
03:21Jarvis key. And we can create it in the default project.
03:25Now I'm going to hit create secret key. Okay. So I copied this key to my clipboard, and don't worry, I'm gonna delete it right after I film this video, but make sure this stays completely safe and private.
03:35And so when I paste my prompt, I'm actually going to run it with my API key. I'll paste it right And so the prompt is relatively simple. Build a Jarvis like desktop AI companion named Jarvi, except let's go ahead and call Jarvi let's make Ricky.
03:49We're gonna make one that's a completely different vibe than the one that I showed you earlier. The goal is to create a desktop AI companion that I can talk to naturally while it does useful work in the background. It should feel like a personal Jarvis, conversational, fast, visual, and able to use tools.
04:06Read the docs on GPT Realtime two. This is going to be desktop application. We're using the GPT Realtime two for live voice conversation, let me interrupt, ask blah, etcetera, etcetera.
04:18Give Ricky a minimal animated face. I'm gonna say high quality animations, but just eyes and mouth, nothing else.
04:29Natural blinking, expressive moods, and speech synced mouth movement. We're gonna keep the UI simple, a small companion window plus an artifact panel for anything visual and structured. Don't worry.
04:41I'm posting this whole prompt in the description if you wanna try it out. When using GPT Real Time two, it is a voice model that can use tools. And so all you have to do is ask Cursor to create and add tools to the agent, and it'll be able to just do all of these really cool things.
04:57And so one of the tools will be web search, and for this, we're gonna be using Exa. We're also going to be generating images. So this voice agent will be able to generate and edit images with GPT.
05:09I also want it to be able to create mermaid charts. This is a coding language that animates or it it renders high quality charts.
05:18And so this is gonna be really cool. I'm just going to copy this entire prompt. Now I'm going to go into cursor.
05:24Now what I'm going to do is I'm going to use GPT 5.5 high, and I'm going to paste this prompt, and I am now going to run it.
05:35So we are running this prompt, and now Cursor is going to build it. So now Cursor is probably going to take ten to fifteen minutes building this app.
05:45I'm gonna go get a little bit of dinner and come back, and we'll see what Cursor cooked up. Alright, guys. So I just came back, and it is now done.
05:54And it created an electron desktop app, and here it is open on my computer. This is Ricky.
06:15This is cool. Can you list out the things that you can do? Let's map out what I can handle for you.
06:19I can search the web, show results, notes, tables, charts, and images in the artifact panel, and keep simple local notes and records. Can also help Wait. Can you just display this in a mermaid diagram in the artifacts panel?
06:31Is that possible? I'll sketch the capabilities as a simple mermaid diagram in the artifact panel. Done.
06:36It did not render correctly. It says parse error on line eight. I'll simplify the node labels and rerender the diagram so it parses cleanly.
06:45So it can create little diagrams. Alright. Very, cool.
06:49This is amazing. Can you search the Internet? Yes.
06:53Give me a specific topic or question. Search the best places to eat in New York.
06:58Off of a few reputable lists and recent roundups that summarize the common standards. I can't access web search right now because of search.
07:05I just realized I forgot to give you the Exa API key. There's a lot I wanna edit here. Hey.
07:11For the past few weeks, we've been helping companies get useful agents set up in their businesses. And these are agents that can be shared with your entire team. So our team can come in, analyze all of your workflows, find where agents can save you time, and be incredibly useful across ads, research, content, reporting, internal docs, whatever it may be, we can locate where agents can help your business.
07:34And then together, we can build shareable agents that everyone in your company can use. These agents will have the right context, the right tools, and most importantly, the right permissions. So your whole team could just use agents where you already work because we can put these agents in Slack, and we can even put them in iMessage or wherever else you communicate as a team.
07:53So if you want help becoming an agent native company, the link is in my description. Reach out there. I'm going to disconnect from Ricky.
08:00So we disconnected. Now I need to make some changes on Cursor.
08:04Okay. So I'm gonna list out a bunch of things that I want you to change right now. The first thing that I want you to change is I want you to fix the mermaid diagrams.
08:11There was a parse error, and I wanna make sure that never happens again. I want you to fix that. The next thing that I want you to fix, I am going to give you my Exa API key, and that's how I want you to make the web search work every single time.
08:25So I'm gonna paste that at the bottom of this prompt. And then for the notes, I noticed that it saves them locally which is awesome. And then overall, I want a design overhaul.
08:34I want the site to have less like components. You know, there's too many like rounded edges on this thing. When I look at this, it's just kind of ugly.
08:43I want it to just have a simpler interface and have the buttons. We don't like those buttons like connect voice display computer use artifacts. Those can all be very simple buttons down at the bottom of the left panel.
08:54Make the face take up like the left half of the screen, like fifty fifty. And then the artifacts are just on the right. Have like a nice line that separates the left and the right.
09:04Okay. Huge prompt. We're making 10 changes here.
09:07Let's see if it does it. Right before I enter this prompt, I'm going to enter my Eksa API key. So here we are.
09:13We're using Eksa. Eksa is just a web search API. The same way we're using an API key to use the GPT real time voice, we're also using one for Exa.
09:22In my opinion, it is the best web search API. You can give any agent the ability to search the web really fast and at a really high quality. Name this project.
09:30Let's call this Ricky. We're gonna create a key And now I'm gonna copy the key. And now I'm gonna go ahead and paste it.
09:37And we are going to run the prompt. Alright. So now let's wait for cursor to finish.
09:42Okay. So it is now done. Let's go ahead and check out ah, our new updated Ricky.
09:49So let's go ahead and I think I can just click this. Hey, Ricky boy. Can you please create a mermaid chart for me explaining how
09:58artificial intelligence works? Okay. I'll sketch simple flow of an AI system and display it as a market chart.
10:05Hey. Can you actually edit it to be a little more detailed and have like little complete sentence subtitles on it? I'll update the chart with clearer steps and short sentence subtitles.
10:14Done. The updated version has short sentence subtitles for each step, so it reads more like a process description. Very, very cool.
10:22I want you to please do a little quick web search for me. Tell me about the latest news in the World Cup, and then explain it to me simply because I don't really understand soccer.
10:32I'll check the latest World Cup news and then break it down in plain language. Here's the simple version. The tournament is in the knockout phase.
10:42I need to change you. Okay. So I need to go to Cursor.
10:46For the web search, I want it to render the web search as just like clean markdown. I want it to stream out and look really cool and futuristic when it when it does that. And right now, you just kind of pull up a bunch of tables.
10:58I want AI to like, whenever I ask for a web search, want it to render out in in text, you know, make it look good. Okay. I'll fire that prompt off.
11:07Hey, one thing I want you to do as well, I want to be able to ask you to show me the menu. So I'm gonna say show me the menu which allow just lists all the things that I can do. Just to keep track of all the things I can do.
11:19And I also want you to put some more time and effort into making the mouth move more realistically based on the different way the the sound comes out, making the animations look better and higher quality. So it is done.
12:05Thanks. The image generation is still running in the background, and it should appear in the artifact panel as soon as it finishes. You're doing a great job, buddy.
12:13Thank you. You got it. Happy to help anytime, buddy.
12:16Hey, can you please show me the menu again? I'll pop open the capability menu in the panel for you. The menu is up in the artifact panel.
12:22Okay. Can you actually enter computer use mode? Sure.
12:26Switching you into computer use mode now so I can control the desktop. You're in computer use mode now. Can you open Codex?
12:39All set and standing by for the next move. What do you wanna do in Codex? Can you type into Codex
12:45to make the script better for the intro? I just want you to type it into the box there. I can do that, but I need your explicit confirmation before typing into an Yeah.
12:54I I give you my explicit confirmation. Okay. I'll type that into Codex now.
13:12It didn't submit. It still didn't When I go to use the computer use mode and I wanted to type something, it asks me for approval every time. Is there a way to turn off that explicit approval mode?
13:22Are you able to do that? If so, do it. Let me see.
13:25I'll look at the intermediaries. I found the exact blocker, computer type text reject sensitive monota. Okay.
13:31Let's see if this works. When I enter computer use mode, I want it to the window to get really small and take up the bottom left. I want it to be like really small in the bottom left and it should just show the little character on the bottom left and the background of the whole app should be translucent.
13:45So it should just be the little character, just a little circle at the bottom with a little button at the top right of it that's brings it back to normal mode. Like computer mode, because I wanna see whatever's on my computer. When I go into computer mode, it should just get really small and it should just be a little character at the bottom left.
14:01Okay. So let's give this a try here. Let's see.
14:04PayPal, how are you? Steady and sharp. Ready anymore.
14:09Need you to enter computer use mode. Got it. Switching mode so I can operate the desktop directly.
14:15Look at it. Alright. Please, can you open up Codex?
14:20I'll open Codex now from the desktop. Nice. Can you please type in a prompt here and say, can you please make this a little shorter and then fire off that prompt for me?
14:29Sure. I'll type that prompt and send it now. Done.
14:36Okay. I just got a really good idea. I want you to be able to generate thumbnails of me using the GPT real time image one, and then I wanna be able to edit images.
14:46And so the way I want this to work is I'm gonna give you an image of me, and that should be used as an image reference. Look up the GBT image two or the the whatever the latest one is. Look up the documentation when I wanna create an image of Basically, I'll ask you to create a thumbnail, and then I wanna be able to edit the image.
15:01So I can say, like, edit it so that it does this. And then you should be able to edit it and change it just with the voice. And I want to be able to generate multiple versions in parallel.
15:11So like what you'll do is when I say generate an image, I want you to generate them in a grid. So like I'm thinking like a three wide grid and it always generates in position one, the top left one. And then when I tell you to generate another one, they should all be numbered.
15:26And so I could say edit number two and make it blah blah blah. And then that one, the newest one will always show up in position one. And so I can continuously edit these images.
15:36This is kind of this continuous process for editing thumbnails. This will be sick if we get it right. Let's see how Cursor does here.
15:44Wait. And I just actually entered a photo of me. I realized it's gonna need a photo of me.
15:49Alright. Let's see if this works. Let's open up our homie here.
16:10a iPhone. The iPhone screen's facing camera, so it's visible.
16:17And it shows a really cool app on the screen. And it looks like a camera, but with like a cool yellow border around it.
16:28Alright. I'm tuning a clean new 16 by nine file now with your exact iPhone set. Cool.
16:33Generate an image of me holding a MacBook except the screen is visible showing a
16:42little blue character with cute white eyes on Number white 14 is being generated now with the iPhone screen facing the camera and the yellow board nice. I'll create a single new thumbnail with that visible MacBook screen and the blue character. Amazing.
16:58I'm opening number 13 so you can see it full screen. Nice. Can you please give him blue sunglasses?
17:06Got it. I'll make a quick edit to that selected thumbnail with blue sunglasses.
17:11Hey, can you please pull up photo 15? Got it. I'll pull up number 15 full screen.
17:15Can you add text here that says cute character?
17:20Got it. I'll add that exact text to the selected thumbnail. Can you please pull up number 14?
17:25Sure. I'll open number 14 so you can view it full screen. Amazing.
17:28Can you please make the background outside like in the forest instead of like the blue to orange gradient? I'll update the background to a forest while keeping everything else consistent. Can you pull up 16?
17:40The thumbnails are still 16 by Pull up 16. Got it. Pulling up number 16, full screen, so you can see it clearly.
17:50brown sunglasses? Nice. I'll create a new version with brown sunglasses using this selected cool.
17:56And, yeah, I want to do a web search real quick. I'll pop open the menu, then you can give me the web search query. Menu is up.
18:03Tell me what you wanna search for. Okay. I want you to please search and tell me the latest on the rumors about Nano Banana.
18:11I just saw Logan Kilpatrick tweeted about it. Tell me what you can say about that. I'll pull together the latest reliable info about those Nano Banana rumors and summarize what's actually supported.
18:22Short version, it's mostly speculation. I want you to show me a mermaid diagram of how our thumbnail tool and like like generating images work. I wanna kinda see a diagram of how that works.
18:32Mermaid chart. Let me sketch a clear flow of the thumbnail and image steps in a simple diagram. Very cool.
18:38Thank you for showing me that. Can we go into computer mode? Switching to computer use mode now.
18:46So we created this Jarvis and we created it in one sitting using cursor with GPT real time. It also has the GPT image model built in and you can just ask it to create stuff.
18:58That's what I'm trying to show you in this video is you can just talk to AI, get it to create your own little custom Jarvis, and you can even make a useful workflow. I actually enjoyed using that thumbnail image generator. It could use a little bit of work, but it's so much fun to create this little Jarvis.
19:14And, yeah, it's just fun to use. And it can control my computer, it can do a lot of things. Hey, Jarvis, go back.
19:19Leave computer mode, actually. Don't don't do it. Got it.
19:22Staying in computer mode? No. Leave computer mode.
19:24Okay. Leave in computer mode now. You're back in display mode.
19:29Good job. Alright. Well, I'm going to take this entire project.
19:33I'm gonna upload it on GitHub, and so you're gonna be able to take the link, the GitHub link below, and get feed it to Cursor and say, clone this, and you could create this exact Jarvis. Or you can look below in the description and get the prompt, and you can create your own version of Jarvis. And, uh, yeah, let me know what you guys create.
19:49I genuinely wanna see what people can create and what types of Jarvises people are creating because you can just create Jarvis. Like, one thing I would love to see is someone create it connect it to their email, and someone would connect it to just some some of their other tools.
20:03I think that would be really cool. And by the way, guys, I'm here in New York City. I have my entire team is here.
20:09Basically, all we've been doing is helping companies implement AI agents into their business. We will help you set up AI agents in your company. We will come into your Slack and we will add agents.
20:23You and you can just switch between the different models. You can also use the open source models if you want. We'll take a look at all of the workflows inside your business, and we can help you create agents.
20:31The most useful agents that we are creating for companies right now are marketing agents, specifically ones that can, like, scrape high quality ads and content from every single platform and then create versions for your company, controlling your ads and running your ads and managing that. And so ClaudeCode, Cursor, and Codecs are all individual tools.
20:52We wanna come in and help you build agents for your entire team, and that means creating skills that anyone in your company Slack can use, and so we will help you set that up. And if that interests you, the link is in the description. Anyway, thank you guys so much for watching.
The video opens mid-demo, not mid-explanation: a live voice exchange with an already-working AI companion that opens menus, changes its mood, and takes control of the screen — before Riley Brown rewinds to show that the whole thing was built in an afternoon using nothing but prompts.
Frameworks
Named ideas worth stealing.
15:07concept
Numbered image-grid editing loop
Generate into a 3-wide grid
Newest image always lands in position 1
Refer to any image by its number
Edit-by-voice targets a specific numbered image
Loop continues indefinitely
A voice-only image generation workflow where every new image is numbered and positioned predictably, letting the user say 'edit number 2, make the background a forest' without ever clicking or pointing.
Steal forany voice-driven or hands-free creative tool needing unambiguous reference to multiple generated assets
08:01concept
Bundle-the-feedback prompt
List every bug and feature request in one prompt
Include bug fixes, new API keys, and design changes together
Run once and review the combined result
Rather than iterating one small change at a time, Riley batches roughly ten distinct fixes and feature requests into a single prompt before re-running Cursor.
Steal forspeeding up any AI-coding-tool workflow by reducing round trips
CTA Breakdown
How they asked for the click.
VERBAL ASK
19:30product
“I'm gonna upload it on GitHub... you can look below in the description and get the prompt, and you can create your own version of Jarvis... if that interests you, the link is in the description.”
Soft dual CTA — free GitHub repo/prompt for viewers to self-serve, plus a separate pitch for Riley's team's paid agent-implementation service woven in mid-video and again at the close.
A 22-minute weekly roundup covering the open-source model that finally passed the vibe check, Codex screen-to-skill recording, the SpaceX/Cursor acquisition, and two Claude updates.
A 20-minute breakdown of the $60B SpaceX–Cursor deal, a hands-on tour of the platform, and a one-prompt playbook for migrating your entire Codex or Claude skill library.
A 22-minute breakdown of Anthropic's Mythos-class Fable 5 model — live demos, a structural argument about building blocks, and a pricing window closing June 22.
A 26-minute breakdown of OpenAI’s Intelligence at Work event: Codex merges with ChatGPT, role-specific agent plugins become startup killers, and Sites turns vibe coding into hosted apps.
A 39-minute live build: one AI agent cloned from a template, one built from scratch, both running 24/7 in Slack and iMessage using the GRASP framework.
A 33-minute field guide to the nine structural shifts in AI agents — what's permanent, what's changing, and the five human foundations that will matter most.
00:22
00:39
00:49
01:11
01:27
01:58
02:14
02:30
02:46
03:02
03:18
03:33
03:49
04:21
04:37
04:53
05:09
05:24
05:56
06:12
06:28
06:44
07:00
07:12
07:28
07:40
07:57
08:19
08:35
08:50
09:12
09:22
09:35
10:10
10:26
10:41
10:57
11:13
11:29
12:01
12:17
12:32
12:48
13:04
13:20
13:36
13:52
14:07
14:39
14:55
15:11
15:27
15:43
15:58
16:14
16:30
17:02
17:14
17:34
17:49
18:05
18:21
18:37
19:09
19:24
19:40
19:56
20:07
20:29
20:46
22:18
Video of the Day20:38
22:24
26:18
39:16
33:38