47:00David Ondrej · Tutorial
100 hours of Hermes Agent lessons in 46 minutes
A 47-minute walkthrough of all seven levels of Hermes Agent — from bare VPS to full MCP back end.
May 6thThe argument in one line.
The developer who wins in the agent era is not the one who writes the most code but the one who provides the best context — because every quality gap between AI outputs is a context gap, not a model gap.
Read if. Skip if.
READ IF YOU ARE…
- A developer or indie founder who actively uses Claude Code or Cursor and wants to know whether Codex is worth switching to.
- Someone building a new SaaS product who wonders whether a traditional browser-based UI is still the right delivery surface.
- A creator or builder who wants to understand how to use public demos and a Twitter audience to de-risk fundraising before launch.
- Anyone who uses AI coding tools daily and wants workflow shortcuts (keyboard macros, model tiering, multi-agent spawning) they can steal immediately.
SKIP IF…
- You want a deep technical tutorial — this is a high-altitude strategic conversation with occasional screen demos, not a step-by-step build.
- You are not yet using AI coding tools at all — the nuance here requires baseline familiarity with Claude Code or Cursor.
The full version, fast.
Pietro Schirano, who spent nine months at Anthropic working on Claude Code and MCP before leaving to build MagicPath, argues that Codex has surpassed Claude Code on two dimensions: a tighter agentic loop and significantly lower token consumption. But the deeper argument in this conversation is structural — software delivered through websites is dying because every person will soon run their own agent with their own context, model, and harness. The practical implications: build less UI, more API; invest in domain-specific knowledge the agent can consume; use brand and community as your moat because features are ephemeral. On the workflow side, Pietro shares the macOS text replacements that expand into full Codex prompts, a model-tiering trick (5.5 high for planning, small model for execution), and multi-agent spawning for parallel tasks.
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 →Who's talking.
00:26guestPietro Schirano
00:00hostDavid Ondrej
Where the time goes.
00:00 – 00:57
01 · Cold open — Pietro's credentials
Host intro, Pietro's Anthropic background, teaser of key claims: left Anthropic, raised money in a week, Codex vs Claude Code.
00:57 – 05:04
02 · Codex live demo — MagicPath inside the agent
Pietro shows Codex's browser/terminal/file panel, demonstrates MagicPath running inside it, explains why software delivery via websites is dying.
05:04 – 06:08
03 · Sponsor — Oxylabs
Real-time web scraping API with MCP server integration.
06:08 – 11:28
04 · Agents eating software — the structural argument
Browser becomes secondary surface. Codex SDK prediction. Agent-first vs browser-first. MagicPath on mobile demo — Pietro builds a David Ondrej website from his phone while walking.
11:28 – 16:58
05 · How to get an unfair advantage
Idea-to-share-to-collaboration speed as the key metric. Three things to optimize: the problem, the context you provide, the easiest distribution surface. Building a legal AI company example.
16:58 – 20:35
06 · Brand and community as the moat
Pietro's Twitter-first brand-building story, the Figma-is-dead tweet, stick to your truth publicly before you build the product.
20:35 – 25:33
07 · Building from zero — show, don't wait
Advice for founders without audiences: just post demos. The Fazza viral moment as proof. Pietro's 2024 Claude speech-to-text demo as a before/after on model capability.
25:33 – 30:04
08 · What makes a great demo
Screen Studio as the alpha tool. Keep it under 60 seconds. Coherent story beats impressive features. The MagicPath 40-second Codex demo that drove 400K views.
30:04 – 32:54
09 · Distribution — Twitter vs YouTube
Twitter for alpha and early movers. YouTube for actual customers and long watch times. Both required; different jobs.
32:54 – 36:37
10 · Startup lifecycle — ship, listen, stay in your niche
Perfect is the enemy of good. One feature at launch. Discord beta testers for MagicPath 2.0. Don't be Steve Jobs — get feedback early.
36:37 – 40:04
11 · Competing at every stage
Claude Design as the big-company competitor. Stay in your unique lane. MagicPath as agent-first, deployed React — not a Figma competitor.
40:04 – 43:52
12 · The only bet you can take
Models are always getting better — that is the one certainty. Build incrementally toward that future. MagicPath roadmap from AI design tool to agent platform.
43:52 – 47:12
13 · How Pietro actually builds — keyboard macros and model tiering
absorb, spawn, PR, bug, plan, explain text replacements. 5.5 high for plan → small model for implementation. Multi-agent spawning for parallel tasks. Codex on mobile via iCloud sync.
47:12 – 50:03
14 · Flipper Zero demo and final action steps
Pietro's Codex-built Flipper Zero game and vibration notification. Three action steps: sharpen context skills, use MagicPath as your canvas, build and ship ideas publicly.
50:03 – 48:03
15 · Sign-off
Mutual appreciation, hint at upcoming OpenAI model, goodbye.
Lines worth screenshotting.
- Codex uses dramatically fewer tokens than Claude Code for equivalent tasks — the benchmark gap is visible on SWE-Bench, not just anecdote.
- The browser is becoming the secondary surface. Products built around 'go to a website and do the thing' are on borrowed time.
- AI quality is a context problem. The same model that produces slop for most users produces beautiful output when given the right context — the model is not the variable.
- Build less UI, more API. The winning startup architecture for the agent era is minimal interface, rich domain context, and an endpoint any agent can call.
- Use your smart model for planning, your cheap model for execution. Prompt the big model to write a junior-developer plan, then switch to the small model to implement it.
- Keyboard text replacements that expand into full Codex prompts are a compounding workflow advantage — and they sync to iPhone via iCloud for mobile Codex.
- Multi-agent spawning in Codex lets you break a task into parallel workstreams that complete in five minutes instead of an hour.
- Brand and community are the only durable moat in AI. Features get cloned; a point of view held publicly for years does not.
- Raising money from a demo tweet is reproducible: leave a job, build in public, make a 60-second Screen Studio video, post on Twitter.
- Atlas (ChatGPT) failed because it was browser-first, agent-second. Agent-first, browser-second is the correct inversion.
- The future of work is less about doing the thing and more about supervising the thing — the skill that compounds is context-setting, not execution.
- A 40-second Screen Studio demo got MagicPath 400,000 views and bent the user growth curve — shorter is almost always better for product demos.
- Figma's file format is a clusterfuck of JSON — that complexity is why agents can't build Figma files but can build equivalent React in 20 lines.
- Perfect is the enemy of good: MagicPath launched with exactly one feature (AI-generated design) and used that constraint to grow.
- Twitter has the alpha crowd; YouTube has the customers. You need both, but for different jobs.
Context is the skill that compounds in the agent era.
WHAT TO LEARN
The gap between a 1x and 100x developer is not which AI tool they use — it is how precisely they provide context to that tool.
02Codex live demo — MagicPath inside the agent
- AI output quality is a context problem, not a model problem — the same model produces slop or excellence depending on what context surrounds the prompt.
- The browser is becoming a secondary surface — products that require users to 'go to a website and do the thing' are on borrowed time.
04Agents eating software — the structural argument
- The future of work is supervising agents, not executing tasks. The skill that matters is knowing when to intervene, what context to inject, and how to evaluate output.
- Agent-first design (agent as primary actor, browser as add-on) produces fundamentally different — and more useful — products than browser-first design.
05How to get an unfair advantage
- Build for agents as the primary user: minimal UI, rich domain context as an API, and a distribution surface that pre-sells the product to humans.
- The three questions that matter: what problem am I solving, what context can I provide the model, and what is the easiest way to surface that product to an agent?
06Brand and community as the moat
- Brand is the only moat that compounds: establish a public point of view before you launch a product, so the product arrives into an audience that already trusts your judgment.
08What makes a great demo
- Raise money before you build: a 60-second demo with a coherent story, posted on Twitter, can generate investor conversations the same week — no deck required.
- Screen Studio's auto-zoom makes any demo look polished without editing — the tool that makes the demo, not just the demo itself, is part of the alpha.
09Distribution — Twitter vs YouTube
- Twitter surfaces alpha; YouTube converts customers — use both, but do not confuse them: Twitter audiences drop off after 60 seconds, YouTube audiences watch seven times longer.
10Startup lifecycle — ship, listen, stay in your niche
- Perfect is the enemy of good: launch with one feature, get 20 beta testers from your community, and iterate — nobody has Steve Jobs's ability to know what people want before they see it.
13How Pietro actually builds — keyboard macros and model tiering
- Keyboard text replacements that expand into full, battle-tested prompts are a compounding advantage: they encode hard-won judgment into muscle memory.
- Use your most capable model for planning and a cheaper model for execution — the expensive model's job is to write a plan clear enough that any model can implement it.
- Multi-agent spawning — breaking a task into parallel workstreams — reduces build time from an hour to minutes for tasks with separable subtasks.
Terms worth knowing.
- Codex
- OpenAI's agentic coding assistant with a built-in browser, terminal, and file system access. Distinguished from Claude Code by its claimed lower token consumption and tighter agentic loop.
- Agentic loop
- The cycle by which an AI agent observes a state, decides on an action, executes it, and observes the result. A tighter loop means fewer wasted steps and more reliable task completion.
- MagicPath
- An AI-native design canvas (magicpath.ai) that generates and deploys real React websites rather than Figma mockups. Designed to be called by coding agents as an external canvas.
- MCP (Model Context Protocol)
- An open protocol that lets AI agents communicate with external tools and services. Pietro worked on MCP at Anthropic before leaving to build MagicPath.
- Text replacement / keyboard macro
- A macOS feature (Settings > Keyboard > Text Replacements) where a short trigger word expands into a full pre-written prompt. Pietro uses these to inject complex Codex instructions with two or three keystrokes.
- Model tiering
- Using a more capable (and expensive) model for the planning phase and a faster, cheaper model for the implementation phase — reducing cost without sacrificing plan quality.
- Screen Studio
- A macOS screen recording app that automatically zooms into cursor movement, making demo videos feel polished and easy to follow without manual editing.
- SWE-Bench
- A benchmark for evaluating AI coding agents on real-world GitHub issues. Used here as shorthand evidence that Codex outperforms Claude Code on measurable tasks.
- Agent-first vs browser-first
- A design philosophy distinction: agent-first products treat the AI as the primary actor and the browser as an optional add-on; browser-first products (like ChatGPT's Atlas) put the browser at the center and add AI as a layer on top.
Things they pointed at.
28:10linkFazza viral demo (voice AI computer control)
Lines you could clip.
01:49
“I haven't touched Claude Code in, like, five months.”
09:27
“The future of work is gonna be less about doing the thing, but more about supervising the thing.”
04:45
“The idea of going to a website to do something, I think it's gonna die.”
30:14
“The absolute alpha is the one tool you wanna use is this one called Screen Studio.”
45:56
“Engineering is a problem of context — coding, debugging. You wanna get very sharp in the way that you provide context to the AI.”
47:02
“I left Anthropic. I built a demo of Magic Path. I think I got, like, a million views on that tweet, and then I raised money in a week after that.”
30:14
“AI slop is because they don't provide the right context.”
Where the conversation goes.
00:57 – 06:08denseCodex vs Claude Code
06:08 – 11:28denseAgents eating software
11:28 – 16:58denseProduct strategy for the agent era
16:58 – 25:33steadyBrand and community building
25:33 – 32:54denseDemo strategy and distribution
32:54 – 40:04steadyStartup growth and competition
40:04 – 43:52steadyLong-term bets and AI trajectory
43:52 – 47:12densePractical workflow — keyboard macros, model tiering
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.
00:00I left Anthropic. I built a demo of Magic Path. I think I got, like, a million views on that tweet, and then I raised money in a week after that.
00:06So you prefer Codex over Cloud Code right now? So I haven't touched Cloud Code in, like, five months. The only thing you can bet is the Mars are getting better.
00:15The future of work is gonna be less about doing the thing, but more about supervising the thing. This is Pietro, the founder of Magic Path. Pietro also used to work at Anthropic,
00:26is very close to the top people at OpenAI, and has access to AI models and agents that me and you don't have access to. In this podcast, we talk about why he stopped using ClothCode, how AI agents are going to eat all software, why he uses codex to do everything, and which AI company he would build today if he had to start from scratch.
00:45This interview is full of alpha. So if you're serious about AI, you have to watch until the end. This is the David Andre podcast.
00:52Enjoy. Right, Pietro. So you mentioned that tools like Codex and CloudCore are going to eat software.
00:57Can we start with you showing how you're currently building with Codex? And then let's jump into more of the what future of the AI agent world looks like.
01:05Yes. So here I have Codex. Right?
01:08And there's a couple tricks that I do when I work with Codex. I'll show you in a sec. But first, wanna show you, like, what I mean by Codex is gonna eat software.
01:16So as you know, Codex has a browser, right, that you can access. You can go to the side panel, and then you can just type your files, your terminal review, and the browser. So what we build, we basically build the functionality so that our product, which is Magic Path, can actually interact with these agents.
01:30What I meant by codec is the way it's softer. This is what I mean because now codecs is one of the biggest and strongest armies.
01:37Right? So when they use it now, are using, like, their coding agent, you know, they already have their their files. They have their design MD.
01:43They have the skills. People now have their way of working. What's missing is that sort of, like, place where the agents and humans can work together.
01:52And so now you can see, like interface, basically. Exactly. Yeah.
01:54So you prefer Codex over Cloud Code right now? Yes. So I haven't touched Cloud Code in, like, five months.
02:00There there's two reasons why I think Codex is better. One, the RNNESS is better. I think that they build a better agentic loop than Cloud Code.
02:08Cloud Code is, like, too bloated, but I think the most important thing is, like, how much more token consumption like, the consumption of Codex, I think, is way better than Cloud Code. You can see, like Yeah. I mean, this on deep SWE.
02:18Like, on the benchmarks, you can see it. You know? Codex is crushing Cloud Code.
02:22And, uh, as you said, for the same task, it uses less tokens. Yeah. I mean, this is if this was OPUS once, it will take forever.
02:28And by the way, right now, I'm, like, on a high. Like so it's, you know, if I use low, of course, it would have been much, much faster. But if this was OPUS, it will take, like you know, it will take forever.
02:37It will cost a lot of money. What about front end, though? Because, like, a lot of people prefer using Opus or even Gemini for front end because OpenAI Yeah.
02:44You know, it's it's great. Large g factors, but I think the front end is still weakness. Well, you can see it's a it's a context problem.
02:51Look how much better. These are made by Codex. Look how much better those designs are.
02:55Right? And the reason is because when we run it through MagiPath, we provide some context
03:00to the model to act as a designer. I wanna touch on more on, like, how it's eating software because not everybody sees into the future like you do. You know?
03:07And you have obviously, like, connections with people who are really on the cutting edge and and beyond. So, like, explain how agents like Codex or Clotco or Hermes or Pie are eating software. Like, do you think, like, SaaS business models will die?
03:19How people should think about starting a business? Yes. That's a great question.
03:23So I think there's two reason why
03:25software is dying and and Codex is eating everything. People now are moving to a direction where they really like to have their own process, and they really like to have their own context. Right?
03:36Think about this. In the past, let's say that you were working in a team. Right?
03:41And you so there was, like, a team. There was, an engineer, a PM, a designer.
03:47Right? The amount of tooling that those people had access to was very well, first of all, there was not a lot of tools, and it was very limited to whatever activity or work this person was doing.
04:00You were an engineer, you were on Versus Code. You were a PM, you're just leaving Google Drive. You were a designer, you were living in Figma.
04:07But now, because of AI agents, not only the roles are collapsing, so a designer can be an engineer, engineer can be a PM, a PM can be engineer.
04:15Everybody's now coming up with their own tooling. Right?
04:19It's like, oh, some people like Cloud Code. Some people like Cursor. Some people like Codex.
04:24And because of that, it's much harder now to say to somebody, this is the tool you have to use because they're so used to know how to run their own agent. And so for me, my only advice if if I was starting this company from scratch roughly a year ago it's it's it was literally, like, a year ago, maybe, like, two days ago.
04:41If I was starting this from scratch again, I would say, well, you know, build this so it can be integrated with Cursel.
04:49It can be integrated with Codex. Can be integrated with CallCode. Because the idea of going to a website to do something, I think it's gonna die.
04:57And so to answer your original question how you can succeed in this environment, the only thing that matters is brand and community.
05:04It's like you really need to have the community that rallies behind you and say, like, this is what I wanna the the platform that I wanna use when it comes to interacting with those agents, basically. By the way, here's the problem with every AI agent you build. It can't see the live web.
05:19It's reasoning over data that's months out of date, which limits its power. Oxylab fixes that. Their web scraper API pulls structured real time data from almost any public site, and it ships with an official MCB server so you can easily connect it to Clothcodes, Kerastore Codex in a matter of minutes.
05:37Now your agent can scrape Amazon listings, Google results, or real estate data just by you telling it in plain English. Oxylabs also have a headless browser built for AI agents. It handles JavaScript heavy apps and geo restricted sites across 175,000,000 residential IPs in more than 190 countries.
05:56That's why PCMag named it the best proxy service of 2026. And what's nice is that Oxylabs gives you 2,000 scraped results completely for free. So go to oxylabs.io/david.
06:08Oh, and when you need to upgrade, use code david for 20% off on any plan. Thank you to Oxylabs for sponsoring this video. I see.
06:16So, basically, um, you you're a big believer that, like, in the future, most of the software will be used by agents, and each person will have their own personal agent that Yes. They know
06:24has the necessary context about them, use they use their favorite model, their favorite harness, whatever. Exactly. And the browser becomes the place where this interaction happens and the but the browser stops being the main surface.
06:38It just becomes the secondary surface into which you communicate to those agents. Like, for example, you know what I think is gonna happen? I actually can bet you billions of dollars on this.
06:47One of the things that's happen is that OpenAI is gonna launch an SDK. Right? Okay.
06:51It's gonna be like, let's call it the codex SDK, where you're gonna have basically like, you know, let let's say like right now for example, magic path for how powerful it is is still limited by the fact that is leaning in the browser.
07:04If you use our own agent and you select something, you can see the agent knows that this is selected. Right? The problem is that when I do that, this doesn't show up as a pill into the the browser here.
07:14Right? And so ideally, what you want is that potentially codex becomes this SDK where I can maybe take like some events on the browser and I can say like, oh, send these events from the browser to the chat.
07:27And I really believe that's what's gonna happen, where you gonna be able to build an application that just talks to codecs easily, essentially. There's also, like, a wave of of
07:38tools that, like, kinda work alongside you. Right?
07:41Whether they watch clicks or your screen and stuff like that. So you're you're saying basically that but for the agent. So if the agent needs to have, like, maximal contact with what you are doing and what you're seeing.
07:53Yes. Yes. The agent is always there, is proactive, knows what you're clicking around, knows where you're on the website, knows all these things.
08:01What's also funny, if you think about, like, Chateapiti Atlas. Right? So Atlas was kinda like this in a way, right, where you had, like, the browser and the agent.
08:11The problem with Atlas was that Atlas was browser first and agent second. This is agent first and browser second.
08:20And because you do that, it allows you to build a much more solid experience when it comes to, like, actually the building of the agent. Right? And the rest is just like, you know, the browser just an add ads on at that point.
08:32And I think the reason why Atlas didn't succeed is because people are like, well, if I have to use a browser, I just use a browser. Browser.
08:38Like like, it's it's faster for me to go to Google or, you know, go to magicbattle.ai than just ask Codex to go to magicbattle.ai. Right?
08:45Sorry. Atlas. But now because Codex, it's the best agent and it's agent first, I'm okay to have an interactive with the what I was when I have to do stuff in the browser, essentially.
08:55So, yeah, basically,
08:56kind of seeing in the future that, like, more and more of the tasks and more more of the work will be done by agents. Yes. Humans will then be providing the initial, like, spark, some of the taste, and, you know so so then the alpha is, okay.
09:09What do the agents need? What do what does the interface look like?
09:14And the the I guess, the easiest way to, like, stream the tokens both to the human and from the human. Yes. Exactly.
09:20So, actually, a a really good point about this too is that that you that you're making is, like, when you've built a product right now, you have to think about the future of work is gonna be less about doing the thing but more about supervising the thing. I believe that as the model gets better, because they are getting better.
09:39We've seen in the past, you know, two, three years, they wanna just keep getting better. The reason or the chances for humans who just go and say, oh, I wanna change this pixel.
09:49I wanna move this way. It's basically gonna go to zero at some point. Right?
09:52And so what you wanna actually provide is the platform for you to supervise the agent versus doing the thing. What's very interesting about this I wanna show you I wanna I wanna see if I can show you this as well.
10:03So here's Codex on my app. Okay? This is this is Codex on my phone.
10:07So check this out. Imagine that this was in my hand, but I'm showing here on the screen so people can see. Yeah.
10:11So look what I can see right now. I can say, hey. Make a website in Magic Path about David Andre, and show it to me.
10:23So imagine I'm, like, walking in the street. Right? Yeah.
10:26So because MagicPath now it's a place for agent, it's in the cloud, any agent can access to it.
10:33Right? And so now it's gonna do but but because this is Codex, it's gonna, like, you know, look up online. It's gonna look up you online.
10:40It's gonna create a new project, and we're gonna be able to see this thing. Right? And so for me, this is the future of work.
10:45When, like, I am on the walk and let's say, like, you and I are working on a website together, you send me a Slack message or you pull, like, a linear ticket and you say, like, hey, Pedro, you know, like, I wanted to change this. Why do I have to go on the desk and open something? I can just look at my phone and I can say, like, hey, pull up the linear ticket that David wrote and and fix this design.
11:04Right? The future is this. It's like, I'm not gonna go into the thing and change it.
11:09I can just ask the agent to change it. Boom. Oh, look.
11:12David Andrew website. I mean, dude, this is magic. And, like, you can imagine, I could have been on a walk.
11:18Like like, I you know, I'm showing this as a screen share, but, like, I could have I could have walked here on, you know Yeah. And, like, go on a coffee shop and I could make this. And then when I come back home, this is in my compass.
11:29Okay. So let's let's think about alpha. Is it like it's a setup?
11:31Right? Like, obviously, we talked about, like, the future of business and stuff, but, like, let's talk about the future of different individuals. Right?
11:37Everybody wants to get an unfair advantage. Everybody wants to be on the cutting edge. Yeah.
11:40So what will differentiate
11:42you from somebody else? Just a better setup? It's how easy you solve in a certain issue.
11:48Right? So for example, the easy the the the issue that we're solving right now, it's like, well, I wanna go from the shorter delta from idea to share to collaboration.
11:57Right? And so the way that people should think about when they make a a new company, a new product, a new idea is, first of all, what is the product I'm solving? And then how can I solve that problem so the agents can take advantage of it?
12:11That's the most important thing. Okay. That's the hard part.
12:13Know? That's where, like, the billion Yeah. Ideas are made.
12:16So, I mean, okay. Like, how would you how would you think about this? Let's say you're not doing magic path.
12:20Right? Let's say you have to start something different Yeah. Today.
12:23How would you think about it? I will think about it as as less UI as possible.
12:30Everything is some sort of API call or service that the agent can access to.
12:38And, ultimately, it's about the data that you want to provide and let the agent work with. Right? Like, let's say you and I were starting, like, a legal company, like, a legal AI company.
12:47I will never do a like, I will never do a product right now. I will never start, like, you know, like, an RV or Legora or the other company, like, where it's, a product that you have to go and upload your document.
12:58Like, why you have to do that? Right? What I will do, I will build a really good harness, so, like, a really good context for the agent to understand how law works, all the different, like, you know, inside of the law and, like, making sure, like, everything's correct.
13:12And then having Codex accessing that knowledge, right, plus my, you know, my PDF, anything I'm working on.
13:20For example, right now, I'm working one of my employees from Ukraine, and I'm I'm sponsoring this o one visa. Right? And so there's a lot of documents.
13:28Right? I would love to have a product where I'm just like, I take all my PDF, and I just talk to Codex, but he also has some really good legal legal understanding so that I can, like, work with that.
13:38Right? Yeah. And so, basically, the the the the the the three things you should care is, what is the problem I'm trying to solve?
13:46What is the context that I'm providing them to them to the model, either in the form of knowledge or documents? And then what's the easiest way to serve that product, which now I believe is, you know, inside of the Codex app or Cursor or, you know, whatever you prefer, basically.
14:02Because this works everywhere. Like, if I open Cursor, works with Cursors too. So make make it easy for agents to use.
14:08Dude, look at this. That's insane. It search you.
14:11So now I have a website magic path. Right? And I can just like real data.
14:15The the metrics are correct. It's the image is correct. Like, you you looked it up online.
14:19And I can also bring this to Figma. That's the crazy thing. I can just, like, copy this to Figma if I want now.
14:24Right? So you're you're not seeing yourself as competing with Figma? No.
14:28Because the clipboard is is blocked on because I I have to enable on on on this browser. But it I'm not competing with Figma because Figma to me is still in the, like, pixel phase moving thing.
14:41You know? It it's like ideation first. Yeah.
14:44Yeah. It's not agent first. And, also, their file system is so complicated.
14:49That's why Figma hasn't adopted any of this stuff because it's, like, it's really hard for agent to build a Figma file because it's a it's a it's a it's a clusterfuck. It's just like like, even like a a super simple sign in card, it's millions of line of codes in a Figma JSON versus, like, you know, 20 line of codes in React or, like, HTML or whatever.
15:07So even, like, the code base
15:09should be, like, simple for AI agents, optimized for agents?
15:13Exactly. So the other thing that we did is, like, this also pull from your repository. Like, I can open, like, a a rep when I can say, like, oh, bring my rep into the canvas.
15:22Show me, like, two, four three, four different variation of this idea and try to explore it, and then boom. Bring it back to the canvas. Right?
15:28Because the agent are, like, are, like, code first, which is really cool. Very nice. You mentioned the biggest leverage is brand and community.
15:34Can you, like, expand on that? Yeah. So similar to you where, you know, you build your own there's, like, a huge persona on YouTube.
15:41Like, I built, like, a a persona on Twitter, and it's, like, you know, in the early days, was showing, like, ideas around, like, how to use AI, how to use, like, different coding agent. I built my own coding agent. I remember, like, one of our first podcast was about cloud engineer, like, because it was, like, one of the very popular early, like, terminal coding agent.
15:57And so I think, like, even before you start thinking about what is the product that you wanna build, it's important to start putting yourself out and be like, okay. This is what I stand for.
16:07The one thing that is interesting about me, if you like, I'll serve you from outside, is, like, I was from the early days. I was like, this is the future of design. Like, I remember I had this tweet that was like, Figma is dead.
16:17There was, like, millions of people saw it. It became, like, a a viral topic in The US. And people call me crazy at the time.
16:24This was, like, this was, like, a year and half ago. People, like, attacked me. I thought I was insane, but I was right.
16:30Like, I knew this was coming. And and so I think you you really need to stick to your truth, basically. And we're like, okay.
16:36This is what I believe is gonna happen. Become the person that people associate when they think about that, and then build that product around, essentially. And, of course, like, you need a good brand.
16:45You need a good design. That's all thing that, you know, plays into it as well, of course. So okay.
16:51I'm gonna ask a selfish question here. Like, for me, if you were launching something new, you would only think about products aligned with my audience. Yes.
16:58You don't wanna get out of your niche. Right? Like, let's say let's say, like, you know, I've been talking about AI and coding, AI and design, AI and prototyping for so long that I think if if tomorrow comes out and I launch, like, I don't know, like, uh, even even AI related.
17:14Let's say I launch, a model training company, right, where it's like, you go and, like, you put your your your weights and, like, you're like like, people will be like, woah. You know? Like, why will I work with this guy?
17:24There's already, like, you know, the file AI and OpenRouter, like, all of this.
17:28Like like, why would I use this product? Because I'm not that's not what people associated with me. They associated me with, like, this type of stuff, basically.
17:36Yeah. So okay. Let's say somebody is, like, starting out.
17:40They don't have a huge brand. How would that change, or would you still, like, advise them to, like, build something in this way?
17:49My advice in that case, if we really started from zero, is still to at least show something. Right?
17:55Let's say you have it, like like, because at the end of the day, like, I think the the the problem that people run into sometimes is, like, they don't just try the thing. Right?
18:03And I know it sounds, like, kinda cliche where it's just like, oh, you have to do the thing. But, honestly, the best advice I can give to people is just, like, just put yourself up. Like, you never know.
18:11Like, I don't know if you saw it. There was this tweet that went viral recently with the guy, Fardza, that did the, like, the voice activation thing.
18:17Yeah. I mean, like, you know, that went, like, super viral. Right?
18:20And it was a good demo. But between a demo and a product, there's, like, a a c in between. Like, there's, like, people that you need to hire.
18:28There's, like, marketing, all this stuff. But at least now, he has a starting point where people can be interested to it.
18:35Right? And it's also interesting, like, how much better the models have gotten.
18:41Right? Because it's, like, it's so easy now to build the experiences like that because actually, it's also an interesting thing to talk about. So just just to show you this, I want I wanna show this because I think this is important for the conversation.
18:52So I actually reshare that tweet from Farsa where, you know, he's showing, like you know, he's, like, basically, like, the demo for people that haven't seen it.
19:01He's just talking to his computer, essentially. Right? And I was and and and I retweeted, I was like, you won't believe how hard was to make this thing in 2024.
19:10So this is me in 2024 talking to Cloud three point five Sonnet. But to do this, you will have to combine so many different things.
19:18Right? I would have to combine Cloud Code with the the text to speech, then the speech to text. And at the time, model can even run tools the way they run it today.
19:28So you have to build your own tooling around having the model doing this stuff. Right? All this to say, now the model is so good that if you have a good idea, it's very easy to just do a demo like in two second with a coding agent and then having people get excited about.
19:44Like like because you could you can clearly tell that he built this with, Cloud Code or Codex because he's running in his terminal. He doesn't have any applications. Right?
19:50And so my advice for you is, do you have a good idea? Just ask Codex to make it. Make it a nice video like this.
19:57You know, show yourself, like, talking about the product you just built. If it's interesting, people notice it. Right?
20:02Like, I don't even think it's because he has a large following. He has, like, around, like, 1,000 followers, this guy. But I think the devil is generally impressive.
20:09Right? He's, like, talking to an AI. He's opening a website.
20:12Like, doing a bunch of stuff. And now if this guy wants to go and raise money, he can probably go and raise for this. It's very easy.
20:18Right? Because he has, like, 3,000,000 people have seen this video. So the most important thing is just, like, you just start.
20:23Let's talk about the demo because you've made a lot of successful demos over the last years. What makes a great demo? That's a great question.
20:30Do you do so good asking questions? It's like my why I really like talking to you. So two things.
20:37One is the tool they're using. So the alpha, the absolute alpha is the one tool you wanna use is this one called Screen Studio.
20:46For people that you don't know about this, this is the tool that everybody uses. And so what this allows you to do when you're recording a video and you're clicking around, it's actually zooming there for you so you can like very easily, you know, show the different part of the the thing that you demo.
21:02And in fact, this guy was actually using that. You can tell he was using that. The second thing is you need to make a video that's very short.
21:10Right? Like, no more than a minute long if you're, trying to impress people. Like, a minute to minute and thirty is is a mouth.
21:16And three, it's like, just make a very coherent story. Right? Like, the reason why the Fazza name will work really well is because, you know, he's like, open Spotify, ask you play a song, roll with the volume, and then he's asking like, oh, can you show me, like, my revenue for my product?
21:29Blah blah blah. So it's a it's a very compelling story of somebody using their computer. So find, like, a simple story to explain, and then boom.
21:38Make that. Like, example, you know, recently, one thing that blew up for me and and Magipath was showing how Magipath works inside of codex, which is basically what I just show you.
21:49Right? Or cursor. And so when I did that that that sort of, like, announcement, literally, what I post was this, and this was the, uh, wait on.
22:01This is not just the video. One second. It was this one.
22:04So this is a forty three second video. Right? Almost half a million people saw it, and all I'm showing is I'm opening the browser.
22:12Going to Magic Path, I'm asking for four different music application, and then Codex builds the music application, and they're actually interactive. They're like I mean, the the audio won't come up in this video, but, like, you can play music in this different, like, music player that he has built.
22:28Right? Forty second using screen studio, very simple idea.
22:33Boom. You can use Magic Palette as an external canvas into into course, into codex, and, you know, half a million people, most 400,000 people saw it.
22:41And, of course, now that made our user base go like
22:44that. Right? Because this is something that people want.
22:47People want a canvas for codex. So, basically, also thinking about the distribution. Right?
22:52Like, as people are building it, how to demo it, how to put it out there.
22:56Exactly. So, like, what advice do you have for people? Because a lot them get stuck in the building stage forever and, you know, they focus on adding features and they cannot leave it to the sides, the distribution.
23:06Yeah. So my advice is, like, perfect is the enemy of good. You kinda have to first solve one problem.
23:12Right? When we first came out with Magic Path, we have one feature, which was just like AI generated design. That's it.
23:18Right? And and the way we used to talk about ourself was like the tool for AI to generate designs.
23:24Right? And that was it. And we did that really well, and and that allows it to grow.
23:28But you can have, like, a million feature because then it's really hard for people to understand what you're doing. The message get diluted.
23:36And when it comes to distribution, like, you really need to understand who your user are. So for example, before we land we launched MagiPad two point o, I went to the in my Discord, and I asked 20 people there to be my beta tester.
23:49And I because I sometimes, like, as a founder too or, like, when you when you build a product, like, you are so into it that you don't see the larger picture. Right?
23:59And so when you have people coming to you and say, oh, well, actually, I like this. I don't like this. This is good.
24:03This is not good. Then it moves you to the right direction. So you do need to put the stuff in front of people because, unfortunately, the problem is, like, not everybody is Steve Jobs.
24:12Right? Like like like, what makes job special was that he could just know what people wanted before having people even to see it. But that's, like, a godlike ability.
24:22Right? And and it's and you only acquire after building product for, like, you know, thirty, forty years that he was doing it. And so you kinda have to have that constant messaging with the people that you think might use your product.
24:35So touch with reality, basically. You know? Like, you can easily deceive yourself that your thing is great, and, you know, you as the builder, the founder, you spend the most time with it.
24:44Right? So you know all the features, you know the user interface, all that stuff. And if you're the only user or maybe couple of your friends, you will never see the rough edges.
24:52Exactly. And and the other thing too, AI is so exciting right now that people would just literally do that for free. Like, the way that I do is, like, I offer Magic Path credits.
25:00Right? But, like, if you're not like, if you're a new founder and you say, like, hey. I have a good AI idea.
25:06I'm looking for a tester. You will get so many response online because AI is so hot right now, and it's been hot for a while, and it's gonna gonna get hotter. And that's just what people want.
25:16Like, if you say, like, hey, guys. I'm testing a new AI tool for writing, you will get an insane response.
25:23Right? And and and and and because it's such of, like, exciting time right now. So you you can find testers so easily.
25:30Just have to tweet. Use it to tweet like what you're doing, basically.
25:34So Twitter is basically the place to go, you're saying Mhmm. For people to both learn about AI and demo it and promote it?
25:40Excited. I feel like Twitter is it's still number one in terms of, like, alpha. Like, things are happening, where people are moving.
25:46Like like, I knew that people were moving were jumping ship from Cloud Code to Corsair to Codex, like, so early on because I I I just see it. I was like, interesting. Like, a lot people are talking about Codex now.
25:57Like Yeah. And now you can see it's just like, it's it's almost flipping. Right?
26:01And so I think Twitter still remains the best place for alpha. But then to learn, I would say that YouTube is still probably the best. And and, actually, YouTube is the one that probably gets you more customer because, yes, Twitter x, whatever you wanna call it, has the most kinda like alpha people, but it's also like a very small sort of like bubble of people that are like AI obsessive.
26:24But like if you actually wanna break into the market and you wanna be successful, you have to put on YouTube Because that's where, like, you know, the average your average customer basically lives.
26:34Yeah. And also, like, people watch like, I can see it from the retention charts. Right?
26:38Like, the same video I upload on Twitter and YouTube, even if it does well on Twitter, people just drop off after, like, thirty seconds, sixty seconds. On YouTube, the watch times are, like, way way longer, like, seven times longer.
26:49Because it's just not a platform to make to watch video, which is why, like, I try to stay, like, under a minute when I post something on Twitter because I know their attention span of people are just, like, so low, um, that I'm just not I'm not gonna do it. Yeah.
27:01So, basically, you think everybody building right now should be willing to just get on camera and just demo their thing? Because, like, a lot of people I'm talking to, they're kinda scared of this. You know?
27:09There's they're not they think they're not social media per people and all that stuff. You think those are just limiting beliefs?
27:15Yeah. I think they're they're just limiting themselves. Like, it really doesn't matter.
27:18Like, you you know how many apps I I made in in in Codex that, like, I could just, like, sell if I want it. And the reason why I don't do it is because, Magic Puff is my is my main, like, interest right now.
27:30But, like, I made so many apps that are, like, actually really cool. Like, I can show you this one actually. It's really interesting.
27:36It's called Oddisale. So what this does is so I can I can write?
27:41Right? It says, like, hello. I'm recording a podcast.
27:47And then I can hit tab, and it's, like, auto complete, like, what I want. Right? Or, like, I can also speak.
27:53Like, I can say, like so it's almost like a mind to text connection.
27:58So check this out. Google Docs would Yeah.
28:01Right? And so it's like, oh, I'm here. I'm here.
28:03You know, I'm I'm with with David. We're recording podcast right now.
28:07And then I can, like, tab autocomplete. Like, how insane is that?
28:13Right? Yeah. And you can just, like, do that.
28:15Right? Or, like imagine, like, I can also say, let's say, like, I wanna write something in between here. I can put a parenthesis, and then I can auto complete, even complete in between sentences.
28:25Right? So, like, imagine, like, if I if I post this online, I'd probably go, like, so viral because such a no nobody's ever done writing things like this.
28:34Right? And, like, there's so many ideas. There's so many opportunities around AI that, like, nobody's using, especially with, like, stuff like GPT real time and text to speech.
28:42Yeah. Like, it's just like an entire new category of product, and you just need to find the thing that you're passionate about and just, like, post it. And people will eventually, they would they would buy it.
28:50Yeah. I mean, you mentioned something great. It's like there's a lot of, you know, SDKs.
28:54There's a lot of AI tools that people aren't using. Like, everybody looks at, you know, Opus. Everybody looks at the g p d 5.5, and they think about, like, LLM based things, which, again, like,
29:05that has been around for, you know, four years now, really working well. That wave of, like, startups, as you could say, like, Perplexity, Curacero are, like, the biggest ones.
29:15But people don't even look at, like, you know, what they could build with Nara Bara two, what they could build with, you know, the GBD real time, like you said. There's, like, gonna be many of companies that just figure out one use case and either get acquired or build a large company.
29:29Yeah. Exactly. And just, like, and just, like, run that user case.
29:32Right? And then because, like, it goes back to the whole idea of community and and brand. Right?
29:36Like, once you get that win win thing, it doesn't matter if you're not OpenAI or OnTropic. Like, if you get, like, you know, like, 5,000,000 ARR, like, everybody wants to buy you.
29:46Right? And so, know, we we you won't believe that they might not, like, buying conversation that we had a magic path.
29:53Right? But, of course, I I wanna grow up this company as much as I can. But the the thing is, like, yeah, like, you you could just find that niche and just run it.
30:02And then eventually, like, if you do it right, you can succeed. Because at the end of the day, AI is just one thing one thing I want people to understand that to make the best product in AI is just a context problem.
30:14You provide the right context, this thing works. When people say, like, AI slop, it's because they don't provide the right context.
30:21Like, you know how many people hate five point five at design? Like, look what we just did. Like, this doesn't even look like 5.5.
30:27This looks like Hoppus. Right? But it's just because I provide the right content.
30:30Like, the interesting thing too here, it generated an image. Right? It went and made his own image here to use, like, GPT generation image to create to put it into the website.
30:39Like, I don't even have to say it to do that. Right? I mean, those models are just so smart right now.
30:44All all it matters is the is the context.
30:46Okay. You mentioned, you know, the the current stage of the startup you are. You know, you're getting offers to get acquired.
30:51How was the different stages? Like, you mentioned this, like, around one year old, a little bit over. Explain to, like, the different problems at each stage, what people need to get right on each stage.
31:01Yeah. So, like, basically, the life of a founder, you are alternating between the best day of your life and the worst day of your life constantly. Right?
31:10So MagicPod is actually an interesting example because when we launched, we were the only product that was doing AI and design, and I was called crazy for it. Right? As I said before, when I was talking about Figma.
31:19Now it's like everybody's fucking doing this thing. Like, everybody. It's like you get, you know, you get Claw Design, which I think it's crazy.
31:28It's like an awful product. Nobody's using it. Right?
31:31But, you know, the normie on the street that never heard about AI and design, he hears Claw Design, of course, he's gonna use it, which is very unfortunate because you get this, like, when a company that big trying to take a name, like, design and make it their thing, it's it's not as good for the startup ecosystem. Right?
31:46Because, you know, we're all trying to compete for attention and and and distribution and all of that. And so I think, like, the one thing that's important is that at any stage of a start up, you always gonna compete for attention and you always gonna have competition. The only thing that matters is that you stick to your to your ideas.
32:05And, like, you stick to what makes you unique. For example, when I was saying, like, I don't see myself competing with Figma, is that ImagiPath, for instance, like, we offer some, like, edits.
32:14Right? And, like, you can change text. You change the color.
32:16You can put the grade and border, like, all the stuff. Right? The thing you cannot do, it's, moving stuff around.
32:21And that's because, like, this thing is deployed. Like, this is like a real React website. It's not like, you know, your Figma JSON file.
32:29Right? And so there's an argument you could be making. It's like, well, why don't you do that, Peter?
32:32Like, why don't you make this editor better? Why don't you invent a new technology that makes the React editing? But the thing is, like, I don't I don't wanna compete with that because Figma spent the past fifteen years building the best editing experience for any kind of this type of experience.
32:46So it's like, I can only lean into the things that may be unique, which is it's fully agentic. It's, uh, something that AI really understands. And, you know, anything is, like, deployed and and lives in in in the browser.
32:59Like, I can just, visit any website that I create. Right? That's what makes us special.
33:03So I think, like, as a founder, as you get to this, like, competition at any stage and and and and just like, you know, all these different competitor that you have, you can just stick for what made you special in the first place, and you just, like, keep keep running that, basically. Okay. So
33:18why don't like, you mentioned you wanna build this as big as possible. How did you make that decision? How would you, like, advise people what to build for?
33:26Right? Because there is two arguments. Obviously, you know, build go as big as possible, as ambitious as possible.
33:32But the counterpoint to that is, like, the future is very uncertain. Right? Like, people can maybe try to see three to six months ahead of time.
33:39We don't know what things will look like past AGI or super intelligence. I don't know if you're a believer in, like, short short timelines, long timelines. But, like, how would you advise to to people, like, okay.
33:48Build something that people know right now is the need for it. Right now is the time for it. Maybe, you know, go for a quick acquisition or try to, like, predict where the future will be going, build something that the agents would use.
34:01Yeah. How would you think about that? Yeah.
34:03That's a great question. So
34:05the basically, the only thing you can do is taking one or two bet. Right?
34:11And one or one or two things that you are, like, so sure is gonna happen. Because as you say, things change so much.
34:18I mean, remember, like, when we first talked about coding agent, now it's just like you know, and you're like, four months ago, nobody would even thought that Anthropic was gonna be the throne as a coding agent. Now everybody's on codex. Right?
34:29So thing changed so so much. So it's very tough to predict necessarily what people are gonna need.
34:36The only thing you can bet is the Mars are are getting better. The morals are constantly getting better and whatever you wanna call it, like super intelligence, post AGI, whatever that's gonna happen, up to that point, the morals are just constantly gonna get better.
34:52So if that's the thing you believe and I think that's the only thing you can believe, that's what you build around that. Right?
34:58And so for instance, like, know, even when I was working on Anthropic and I was working on, like, on search and cloud code and MCP, all those stuff, I think I stood down inside the company because I was building almost for, like, you know, Office seven.
35:11Like, I was building the tools and application thinking that those models were already intelligent. And and so what what that what what's very interesting about that is that how you build this application and experience, you start to see that the models are not there quite yet, but they are good at doing certain things.
35:27Right? And so the only thing you can do right now is saying, well, I believe this is the future of design in my case.
35:33And I know that's where we're gonna go because the models are just gonna get better. And I can only build toward their future incrementally.
35:41Right? So Magipath initially was a tool where you can just do the design. Now it's a platform for AV agent.
35:46And I know that, you know, six months from now, this conversation would be very different where instead of me asking Codex what to do, what probably is gonna happen is if you and I were repeating this in six months from now, you and I are talking and Codex is just making stuff for us. It's a proactive agent.
36:01Right? It's it's not the agent that you ask a question. It's the agent that understands what's happening and work for you?
36:08Right? And so that's the only bet you can take.
36:11Is AI is getting better? Pick up a subset of human experience that you think AI is gonna be amazing at and just build toward that, essentially.
36:23Yeah. I I agree completely.
36:25Can you show us, like, how you actually build software? Right? Like, do you use the Codex app?
36:31Show show us the work. So for for me, you know, so everything that we're doing design wise, every design thing is building MagicPath, and then we use, like, you know, whatever we build in MagicPath, we send it back to the agent to implement into our repository. But, generally, when I, if I wanna start something new or I wanna try a new idea right?
36:48Let's say, like, I'm in I'm in Codex, I have all those, like, quick text replacement in my keyboard.
36:54So, like, there's things that I can do. So for example, if I if I say absorb, it just writes this whole thing, which is, like, absorb this code deeply, you know, solve all the feature in the code.
37:04So that I start that. Say I'm starting with a new repository, I select that, and I start.
37:09Or then I have this one is my favorite, which is all the time, which is is that? This is I'm asking basically to explain this thing that you're working on as you were explained for a junior developer and then write a plan for the junior developer to build that.
37:26And the reason why I like is because if I use plan mode in in Codex let's say I was using, like I think it's, like, slash plan, I think it is. So if I do this and then I wanna change the model after the plan, it doesn't allow me to do that. Because if you select let's say you select, like, 5.5 high in plan mode, then you can change to low mode to implement that plan.
37:47Right? Because the thing is, like, the the smartest model you need is the one that needs to figure out the problem. Then you can just use a smaller model to fix it.
37:55And so the way I like this is because if I ask, you know, 5.5 high to come up with a plan for a junior developer, then I go here, I change it to low, and then I get that one actually implementing the plan. And then another one that I have that I love, I say, if I spell this, I basically told it to behave as an agent.
38:14So I was like, okay. Remember, you're an agent. You decompose the query, blah blah blah blah blah.
38:19You do all of those things. And what's funny, this is basically goal mode. Right?
38:23Like, so today, you don't even I don't I don't I I use this rarely. Like, oh, you can just do, like, slash goal. Right?
38:28And you just you can just write your goal. Another one that I love, if I do spawn, it asks to spawn multiple agents.
38:35Actually, this is, like, really cool. Right? Because let's say I say, make a few images using your image skills of, I don't know, electric electric Ferrari since this was, like, everybody was talking about this.
38:51Yeah. So if I keep my spawn multiple agent prompt, right, now it's actually gonna do that.
38:59It's gonna ask multiple agent to make images. And you can imagine, you can do the same thing where you're coding. You can say like, oh, I'm building this thing, spawn multiple agents.
39:08Right? And and what's awesome about this is that you build the thing in, like, five minutes instead of, like, you know, an hour or whatever. Like, whatever it will take you to build that thing.
39:16So I basically have all these commands. Right?
39:20Or, like, for example, this is my favorite tool. If I if I write PR, it says push a PR towards that branch, you know, with the things I'm working on, or this one is even crazier.
39:30So for example, all times, my my designer my sorry. My developer will work on a branch, and then I wanna take that branch and make some changes and maybe push it into main, like, which is, like, the the main product. But so what I have here, it's like this prompt, what this does is say, okay.
39:45Take this branch that I'm working on. Make sure you pull from the main most recent branch.
39:54Check all everything that matches, cherry pick the things that, you know, are working, and then implement that. So it's like it's such a it's almost like my brain in a way is connected with with, um, with Codex because it's just like I have all those quick things that I can write.
40:09Or if I if I write, like, explain, it asks me like, a GitHub repo? Like, this would go viral for sure.
40:14No. I I know. I I should probably do this.
40:16No. It just it just text replacement. I I just use, like, keyboard.
40:19Actually, you know what's crazy? You can use codex to build this text for business for you.
40:23Like, that's what I do too. I ask codex, like, hey. Go in my text replacement.
40:26Can you add me a few new ones? And it just does it for you, which is kinda crazy.
40:31Yeah. I think you should just, like, take ten minutes, throw this in a GitHub repo, name it Pietro Skills or whatever, and just, you know, release it to be fair. Because, like, there's so many GitHub repos that just, like, get, like, 30 k, 40 k stars in a less than a month.
40:43That's like scale. Yeah. That's just that's just those people's workflows.
40:48Right? Like, it it for us, maybe it doesn't seem as valuable because, we do this shit every day. But for people who don't, this is the gold.
40:55This is huge. Yeah. This has changed my life.
40:57Like, I'm so fast at developing and and and designing all this stuff because it's like, wow. This looks so much better than the Yeah.
41:05It's crazy. A fucking AM model is is able to do better design work. Like, that's like, that's the electric Ferrari I want.
41:11This is 600,000. You know? Not not that shit.
41:14Yeah. The the Lucia fucking sucks, dude. Like, this is this is so much better.
41:19Oh, because I'm zoomed out. That's why. Oh my god, dude.
41:22So funny. Yeah. No.
41:24It's it's it's it's super cool. I do also have a lot of skills. Like like like, that's the other thing.
41:29You don't even need skills when you have stuff like this. Like, I can say, like, the bug, and you've write these, like, really good things. Like, review the recent code that we've wrote, look for potential bug, regression, edge case, side effects, and then list all the risk.
41:42Like like like, why would I use skills? Like, like, this is busy what that's doing.
41:46Right? And I like, you and, of course, like, you can write, like, some really long line, like, one, like But you can do, like, slash.
41:51Right? Like, with skills, you can invoke it with slash. Yeah.
41:54Exactly. But for me, it's, like, know, it's so much faster to just have those, like, keyboard. Because then the cool thing, if you're using a MacBook and an iPhone, then I also have this thing on my iPhone.
42:02So if the same, like, you know, it's it's saved on my iPhone too. So when I'm on Codex mobile, I can do the same thing.
42:08Right? So I don't need to have oh, I mean, skills, I guess, are connected, but, yeah.
42:13It's just a it's just a better way, basically, to to to work. Yeah. So, basically, anything you find yourself repeating, say, like, what, more than once a day, you turn it into these keyword replacements?
42:22So the way you do it, you go to settings, and then I do, like, yeah, text replacement. It's called and then the whole year.
42:34See? But truly, you can say you cannot you can ask codecs to make new ones. Like like, look.
42:39I can literally say, like, this is crazy. Like like like, look in my text replacements for the keyboard and suggest a few new ones.
42:54Yeah. This is, like, a lot of alpha as well, like, realizing how how much control it has over your computer. Like, for example, now I don't even use folders.
43:01If I'm, like, doing something in a markdown file, say, like, open this and find it for me. And it's not, like, clicking seven folders deep.
43:08Yes. Yeah. I mean, that's literally, like, the the the the future.
43:12Right? This is, like this is the best computer. Like, look at this.
43:14You come up and be like, oh, I like Andoff. Right? Like, okay.
43:18Cool. Add Andoff.
43:21Right? Now he's gonna add that in my text replacement. I mean, that's fucking insane.
43:24Right? Like, it's like Yeah. It's crazy.
43:26Right? Yeah. Also, like, people people don't know they see, like, okay, Cloth code or Codex.
43:32Right? And instantly think about coding. But, like, it can do anything in the computer.
43:35Right? Like Anything. Organizing your files, finding the largest files, optimizing this space.
43:39One of my friends was had, like, a old computer from his dad, like, 15 years old. He was lagging and shit. He just used, I don't know, cloth, cloth, codecs.
43:46He said, like, Meg is faster. It ran for, like, five, ten minutes, you know, deleted some files, some apps, optimized processes, and it was, like, way faster than before.
43:53No. It's insane. Yeah.
43:54You you can fix your computer. You can fix your Wi Fi. Can actually do it can download anything.
43:58Dude, it's insane. Actually, let me show this. This is really cool.
44:00Give me a second. I'm gonna show this out there. So I have this device here.
44:04It's called a flipper, although the battery is dead. This is was gonna show you. This is actually really cool.
44:08So here I have it's this is called a flipper. So this is basically like a way to do security control on network.
44:17You can also, like, create, like, cards. Like, you can, like, you know, duplicate cards. And I was like, you know, like, I don't know the coding language of this thing.
44:24It's a device. Like, what the fuck do I know? Right?
44:26Yeah. Yeah. So I connected to Codex, and I was just like, can you just make a game for this thing?
44:31Dude, it made me, like, this, like, double like, uh, like, doodle jump. I don't know if you can see.
44:37And it's like see? I'm like, I I can actually play this game.
44:41Right? But the crazy thing, he made the screen vertical. He was because he probably thought, okay.
44:46Well, control it. Like, if figure out or, like, I can use as a I can also use as a music player now. Or the only interesting thing that I build, when Codex send me a message, this thing says that vibrate.
44:57So I know that Codex has finished to to send a message, which is like how cool like like, I I don't know how to develop for the external device. So like well, like, he hacked this this, like, little e reader so he can make now he made me an app for for this thing.
45:12Like, it's it's insane. We're we're really we're really, like, in the in the most insane era of technology and humanity. It's just, like, so exciting.
45:21So, basically, acquiring as much hardware as possible, improving your AI setup as much as possible so you can smoothly turn your ideas into reality.
45:31Yes. Exactly. There's so much such a such a fun time to act and, like, you know, do new things.
45:36It's, like, just, like, amazing. Yeah. Okay.
45:39Let's leave people off with a clear action step. If you were somebody who just, like, has, you know, the default codecs or default cloth code, what's, like, one or two things you would do today to improve your setup?
45:49Yes. So first of all, get, uh, any skills that makes you do what I basically show you, like, just, like, improve your workflow. At the end of the day, engineering is a problem of context, coding, debugging.
46:02Right? So what you wanna do, you wanna get very sharp in the way that you provide context to the AI, but then also how the AI is building the code and then also how you're looking for bugs in the code.
46:12Right? So you have a you have to have at least solid test sort of pipeline.
46:16That'd be the the the first thing. Second thing, if you need, like, a a canvas to, you know, design and build actual functional apps, of course, I'm biased here, but MagicPack is amazing. Yep.
46:27I didn't show you how to install it, but it's basically a skills. It's a maybe we can we can add it to the YouTube description, but you just install the MagicPath skills, and then all those agent can Or just tell our agent to browse the web and install MagicPath. Yeah.
46:41That'd yeah. Exactly. But it it need it it needs the skill, basically, right, to to know how to communicate with MagicPath.
46:47And then the third thing would be like, yeah, just build as many ideas as possible. Don't be shy.
46:54Like, nobody's gonna judge you if you get three likes on Twitter. Like, that's I I I I like, in the early days, I will get two likes on Twitter. It's like like, it happens.
47:03Like, when you start, it's like normal. And then, you know, eventually, you build something that people are gonna be like, oh, okay. This is cool.
47:09And then potentially, like happened to me, that's how you raise money. Like, it's like, I left Anthropic. I built a demo of Magic Path.
47:16I shared a tweet. I think I got, like, a million views on that tweet. It was like, here's how you design in the age of AI.
47:21And then I raised money in a week after that. So That looks like the TLDR.
47:25Yeah. Like Of course, like, I I I you know, I'm very grateful. I know I was at the right time, at the right place.
47:31Yeah. Lots of work, great taste, you know, experimenting. Yeah.
47:34There's, like, years of work. It's not gonna happen overnight, but at the same time, you kinda if I didn't put that out, I wouldn't do this. Yeah.
47:41Know? So yeah. Yeah.
47:43Alright, man. Appreciate for your time. Lots of amazing advice.
47:46I took shit ton of notes myself. And I I I hopefully, next time, it's not a year until I have you on.
47:51Yeah. Yeah. Yeah.
47:52Yeah. We'll we'll we'll chat more once the new OpenAI model comes up. Alright, man.
47:56Maybe there's a hint that it might be sooner than people think. Alright, Pietro.
48:00Have great day.
The Hook
The bait, then the rug-pull.
Pietro Schirano opens with three words — "I left Anthropic" — and the rest of the episode explains why that decision was rational. As the engineer who worked inside the Claude Code and MCP teams, he now uses Codex exclusively and hasn't opened Claude Code in five months. What follows is 48 minutes of builder-to-builder alpha on agents, context, and the death of software as we know it.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
