15:40Mark Kashef · Tutorial
These Mods Take Claude Code to Another Level
How reverse engineering Codex via MCP and automated sub-agent routing turns Claude Code into an autonomous desktop operator.
October 5thThe argument in one line.
Frontier models can plan, program and debug physical hardware end to end, so the real test of a new model is a real-world device, not another vibe-coded app.
Read if. Skip if.
READ IF YOU ARE…
- You build with Claude Code and want to see what a frontier model can do beyond web apps, games and demo sites.
- You're building a voice agent and want a concrete setup for voice cloning, wake words and expressive speech.
- You have a hardware project in mind, such as a robot, a smart device or a 3D printer, and want a prompt structure to hand to your AI agent.
SKIP IF…
- You want a full step-by-step course: the creator says this could have been a multi-hour tutorial and only shows the outline.
- You're looking for a business or revenue angle: he says outright this wasn't about generating business value.
The full version, fast.
Most new-model demos show a game, a website or an app. Mark Kashef instead buys a $99 M5Stack desk robot, wipes it, and has Claude Fable 5.2 research the device, back it up, and rebuild its firmware and software. The result, Lil Token, checks his email and calendar, researches with Firecrawl, connects to hundreds of apps through Composio, and speaks in a Fish Audio clone of his voice. He drives the build with an eight-part prompt, uses synthetic audio to simulate over 30 test conversations, and moves from Fable to Opus 5.5 for later iterations. Twenty versions later, the kit is open source.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:10
01 · Meet my AI desk robot
Cold open: the robot answers a calendar question in the creator's cloned voice.
00:10 – 01:21
02 · Taking Claude beyond apps
Why typical model demos are boring, and the plan to test Fable 5.2 on a real $99 robot, using Fable first and Opus 5.5 to iterate.
01:21 – 02:49
03 · What Lil Token can do
Live demo: Firecrawl research, Google calendar and email, Composio apps, touch reactions, and a camera check when he sits it back down.
02:49 – 04:00
04 · Starting with a $99 M5Stack robot
How Claude shortlisted robots, why a full wipe and backup came first, and the hardware the device offers.
04:00 – 07:22
05 · The eight-part build prompt
Research, component breakdown, backup, goal engineering, wake word, response style, restrictions and troubleshooting, including the synthetic-audio test loop.
07:22 – 08:07
06 · Cloning my voice with Fish Audio
A voice clone from 15 YouTube videos, plus bracketed expressiveness directions on every pre-canned line.
08:07 – 09:21
07 · Everything that went wrong
Latency, missed wake words, runaway dancing and misfiring tools, fixed by tracing sessions in test mode.
09:21 – 10:00
08 · How a voice request works
The path from speech to Claude to Fish Audio and back, with a T indicator while it thinks.
10:00 – 10:40
09 · Robot faces and music
Faces made with the Codex CLI mapped to emotions, plus the hard problem of classifying music genres to dance to.
10:40 – 11:50
10 · Why I built this and the kit
A goal of showing frontier models beyond vibe coding, a 3D printer on the way, and the free open-source build kit.
Lines worth screenshotting.
- Most AI model demos show a game, a website or an app, but frontier models can also research, program and debug physical hardware.
- Before wiping a device, have the model read the manual and product page, then back up the original firmware so you know what you can and can't change.
- Let the expensive model plan and build the first versions, then hand iteration to a cheaper one: Fable for the first few versions, Opus 5.5 for the rest of the 20.
- A voice agent can be tested without you in the loop: the Mac played synthetic audio at the robot and Claude simulated over 30 conversations while watching for errors.
- Bracketed direction tags in a text-to-speech prompt, like acting startled or excited, make the same line sound noticeably more expressive.
- Cloning a voice took 15 prior YouTube videos fed straight into Fish Audio.
- Hardcoding safety limits in the prompt, such as drafting email replies but never sending, protects you from hallucinations and bugs in an agent with real inbox access.
- Tool calling that misfires or never fires is the biggest source of bugs, and traces from test mode let the model find which requested tools never ran.
- The robot's own motor noise drowned out the music it was trying to dance to, so stop words had to be taught and tuned separately.
- Putting traceability into the plan from the start, with screenshots every second, gives the model evidence to debug against instead of guesses.
Frontier models can run real hardware projects
WHAT TO LEARN
A structured prompt, a safe backup, and a model that can trace its own failures are enough to hand a physical device to Claude.
- Before touching a device, have the model research the exact make and model, read the manual and forums, and back up the original firmware so you know what's changeable.
- Structure the build prompt in parts: research, components, backup, goals, wake word, response style, restrictions and troubleshooting, so nothing gets left to guesswork.
- Use the strongest model to plan and build the first versions, then switch to a cheaper one to iterate, as the creator did with Fable then Opus 5.5.
- Bake safety limits into the prompt up front. An agent with real inbox access can draft replies but should be told it can never send them.
- Test voice agents without sitting there yourself: have the computer play synthetic audio at the device and let Claude read the traces of over 30 simulated conversations.
- Most bugs come from latency, missed wake words and tools that misfire or never fire, so log every session and have the model compare requested tools against fired ones.
- Expressive speech comes from direction tags, so annotate each pre-canned line with how it should be delivered instead of leaving the voice flat.
Terms worth knowing.
- Claude SDK
- A bridge that lets your Claude subscription drive other software and hardware, which Kashef uses to run the robot's brain on Opus 5.5 for personal use.
- Firecrawl
- A web research and scraping tool the bot uses through MCP to look things up quickly. A flame icon on the robot's face shows when it is in use.
- Composio
- A free library that connects an agent to hundreds of existing apps through a single integration, so the bot can act inside tools Kashef already uses.
- Fish Audio
- A text-to-speech and voice cloning API. Its newer Drama 3 model supports bracketed directions that control tone and emotion.
- Wake word
- A spoken phrase that activates a voice assistant. Here it's 'hey, little token', after which the bot listens for around 10 to 15 seconds.
- Goal engineering
- Spelling out the exact dream outcome for the model before it starts, including which model, services and behaviors the finished system should have.
Things they pointed at.
03:10productM5Stack StackChan desk robot
00:15productClaude Fable 5.2
01:10productOpus 5.5
02:10toolFirecrawl
04:40toolFish Audio Drama 3
Lines you could clip.
00:12
“The average video you'll see demoing a brand new AI model always shows the same three things. A game you'll never play, a website you'll literally never push live, and an app you'll forget about by tomorrow.”
10:55
“You can use these frontier models for a variety of things outside of just vibe coding and building software.”
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.
Hey Lil Token, what's on my calendar today? Just one thing today. The Lil Token demo shoot at 9 .30 tonight, which is happening right now.
Hey, I'm airborne. The average video you'll see demoing a brand new AI model always shows the same three things. A game you'll never play, a website you'll literally never push live, and an app you'll forget about by tomorrow.
But models like Fable 5 .2 can work on real things in real life. Anthropic even launched it with a small snippet teaser video of Fable designing hardware. So I wanted to put it to the test.
Now I'm not a hardware person and I've never programmed a robot from scratch. So I bought this $99 one off the shelf and I plugged it into my computer so Claude could take over and build a personal assistant out of it. Now it can do pretty much anything from answering my questions to research for me.
It can check on my email and calendar and use hundreds of applications all with a single voice command. And most importantly, it can talk back to me in a clone of my own voice. It even knows when I pet it, when the music's on.
And it even dresses culturally appropriately for whatever's playing. Now, it took 20 versions to get here, and I didn't use Fable for every one of them. Even though the usage of this model is way better than Fable 5 .1, it's still red hot.
So I used it primarily to plan and create the first few versions, then I used Opus 5 .5 to iterate until we got to the final version. So in this video, I'm going to show you what it can do and how I used Fable to set it up. And even if you don't want to build a robot personal assistant, you'll get a glimpse of how powerful these frontier models are at manipulating hardware instead of just another vibe -quoted app.
Let's get into it. Alright, so this here is my friend Lil Token. Now he weighs as much as my phone, and I'm gonna show you what he can do.
I'm on camera, sitting on a desk with the best view in Toronto, and my boss is about to make me famous. Big week. Anthropic dropped Claude Fable 5 .2 and Mythos 5 .2 on Monday.
And I didn't hardcode that, by the way. It comes up with all of those lines on its own because it's using my Claude subscription. And it can also go and look up things for me.
And it's designed to show me exactly what service it's using for a particular request. So when you see that little flame on its forehead, it's using Firecrawl for research. Whereas if it's checking my calendar or my email, you'll see that Google icon.
If it's using one of the hundreds of apps it can connect to, it's going to use the Composio. .dev free library that you can hook up to to give it access to tons of tools. And this is my real inbox, not some simulated version of it.
It can read it and draft replies, but it's programmed to not send it on its own. Now the part I didn't expect was Fable even figured out how to manipulate the software so the bot reacts to touch. It also knows when I sit it back down on my desk, and every 20 seconds it takes a quick photo to see if I'm back after being away for a while that it keeps track of on the back end, and then it welcomes me and tells me what's coming up next in my day.
And the recordings and camera check stay on my Mac It even runs a few of my models locally for small admin tasks.
Now, a few weeks ago, I could still pull this off with Opus, but it would take a lot more planning and iteration. And even with Fable, it still broke in a variety of very funny ways. At some point, my beautiful little token had green bands all over his face, so we had to fix everything.
So let me show you what this process looked like straight out of the box. Now, I originally bought this desk robot from a company that's called M5 Stack. And I'm not affiliated with these guys whatsoever.
I just had Claude go and research, and I asked it, go and find the top three. best choices for personal robots that you know that if I hooked it up to the computer, you'd be able to manipulate the firmware, software, everything. Then it came back with a few of them, and I picked this one because it was the perfect sweet spot in terms of price and functionality.
So out of the box, this bot comes with its own phone app, remote control, and smart home integration. But again, I didn't want all of my conversations, data, and camera snapshots being uploaded to some random cloud. So I had to be able to completely wipe it and take stock of every feature and element that it had.
So like you can see here, it has... Things like a touchscreen, a camera, two microphones, speakers, motion sensors, lights, and even a Wi -Fi chip. So all the parts we needed were here, and all I needed Fable to do was rebuild all the software and firmware.
So all I had to do was just plug it in, give it access to the product page so it could read all the specifications, as well as the manual. And by doing that, Fable was able to tell me exactly what we could change and what we couldn't change based on everything that I could see. And before we wiped anything, it made a full backup of the system just in case something goes wrong.
So let me break down the eight different parts of the prompt that I sent to Fable so it could start to spec exactly how it would execute the entire project. And again, if you have a hardware use case that isn't a personal bot, you could still apply what I'm about to show you. So step one, I say use Firecrawl to fully search the exact make and model of this device, read the maker's docs, the product page, and the open source firmware projects for it, and what owners are saying in forums.
Then this is the part where I ask it to break down every single component of the app so I can push it to understand how all the components come together. And then we ask it to make a full backup of the software and firmware in case we mess anything up. And in this phase, I also downloaded the mobile app and then I gave access to computer use to use my iPhone mirroring to go and click around and find my Wi -Fi and hook it up so I could then disconnect from the app and fully work between the app, the computer and my Wi -Fi.
And then we get to the goal engineering part where we outline exactly what the dream outcome is. So I told it to swap the default language model it's running on, which is Quinn with Opus 5 .5 using my cloud subscription. And to pull this off, I told it to use the cloud SDK, which is basically a bridge between your subscription and any form of software and hardware that you want to use it on.
As long as you're using it for personal use cases, you should be more than fine with the terms of service, but obviously do your own due diligence. And then I steer it on how the bot should respond to the clone of my voice using the fish audio. API because they have a newer service that's called the Drama 3 model, which is meant to be a lot more emotive and expressive.
So all the intonations that I hope comes through this video, I tried to basically replicate with my personal agent. And to make it that much more expressive, you can add bracketed directions that tell it when to act startled or extra happy or excited. And it will push the API to create a slightly more emotive response.
And then to make any form of research or scraping really quick, I'll use fire crawl via the MCP Gmail via the Gmail CLI. So I can also access things like Google forms and calendars. And then, like I said, you can use something called composio .dev completely for free to hook up to hundreds of apps that you already use.
Now this last part here is. about all the different faces it has, which I'm gonna dedicate a whole section to in a bit. Then this entire section goes through what the wake up word is, which is, hey, little token.
And as soon as you say that, similar to saying, hey, Siri, it'll activate and ask you exactly what you're looking for. And it's designed to listen for around 10 to 15 seconds, just in case you take a little bit longer to finish your sentence. Then part five breaks down how succinct and concise the responses should be and any nuances related to that.
And part six is where I tell Fable exactly. all the restrictions that it should bake in into the bot so it can't send emails on my behalf, just in case there's any form of hallucinations or bugs. Now part seven was all about troubleshooting, where I asked Fable to create a combination of different ways that it could track and trace all the errors and issues that I could see.
So if I was running the bot and I was testing certain functionalities or its ability to listen to music and whether it would become culturally appropriate to that piece of music, it would take screenshots of every second to trace it and see, is this matching what we would expect? Now, this last part has something that you can apply to a variety of scenarios, especially if you're building or using voice agents.
Instead of me having to stay stuck at home and constantly talk to my bot to take it from complete dummy to something smart, I had my Mac use its speakers with synthetic audio to talk to the bot and have full conversations where Claude would just listen in between the middle and see all the errors it had. So hours later, it had simulated over 30 conversations without me being in the loop.
Now, when it comes to speaking to a little token, this is how simple the voice system is. All that happens is I speak in plain English. Then that gets translated through the fish audio API.
It goes and pulls the ID from my clone voice and it spits it back out with the response. And I was able to easily clone my voice by feeding 15 prior videos on YouTube and just throwing it straight in there. Then I was able to ask Fable to break down.
based on all the API documentation, how do we add these specific annotations right here that tell exactly how to act and what to express and when. So for all the pre -canned responses, Claude was able to add a direction for exactly how they should be communicated. This is one of those lines without it.
Hey, easy, I'm getting dizzy. And this is the exact same line with more expressiveness. Hey, easy, I'm getting dizzy.
Now, even with Fable's robust plan, this wall of shame is a small snippet as to every single thing that went wrong. And to spare you from having to read this, because I'll make all this available to you, the main things were the latency of the bot, its non -responsiveness to touch, to actually speaking, a very long lag time between when I say the thing and when it responds, let alone its ability to distinguish whether I'm talking about a tool or whether I wanted to use a tool.
Also, when I'm not on YouTube and I'm just working away, I have a tendency to mumble. So when I mumbled, little toe... it wouldn't pick that up.
So I had to keep training it on the different ways that I could call it, so I could call it from across the room and it would still respond. Then there's the beautiful dancing. Its own motors were so loud that it could hear the music and start dancing to it, but it would never stop because its motor would drown out all the additional sound.
And I had to even teach it stop words like stop dancing, which it would also tend to mishear. All the other problems really came from tool calling and tool misfiring. So because they had all the traces that Fable or in this case, eventually Opus could analyze, it could keep iterating every single time that I had a session.
And to make this easier, I would have testing periods where it would be running in test mode and it would keep pulling, waiting for my entire conversation to run, to see which tools were asked for that didn't actually fire and vice versa. So if we map out a single request, like me asking what's on my calendar, that will then get transcribed by Claude, which will then pass off to Phish Audio, which will then pass back the response that's basically parroted through my robot.
And while it buffers through a request, it will show a capital T at the top right to show me that it's thinking. So then the overall system looks like this, where you have Claude that's always on. And I also have the system linked so that if Claude is done running or it's waiting for me on my Mac, it can tell the bot and it will announce it to me.
And on top of that, it has access to things like Slack, Stripe, and a series of CLIs and MCPs. It even has its own second brain that lives in the cloud that keeps the most common requests on file so it can respond faster. And because Fable rebuilt the entire firmware from the ground up, we had to program every single screen for every reaction and we use the codec CLI to create all the images that would be assigned to a specific reaction or emotion.
So in just picking up the bot, it would have to correlate the emotional state with the physical state and exactly what reaction it would need to say next. And when it came to music, the easy part was just getting it to realize that music was playing. The hard part was classifying what type of music it was listening to so it could categorize and dance accordingly.
So Fable trained it on different audio waves that were playing and what the most common patterns looked like if you were playing things like jazz or classical music. or Middle Eastern music. Now, again, for me, this wasn't necessarily about generating business value, even though it's actually nice to have a little assistant that tells me exactly what my day is without me even picking up my phone.
My core goal with experimenting with this and walking you through the process in this video is just to show you and show myself that you can use these frontier models for a variety of things outside of just vibe coding and building software. I'm only scratching the surface here. And if it's something that you can plug in or connect to your laptop, it's something that Fable will be able to help you with.
I even have a 3D printer on its way. So I'm going to be using Fable to help me figure out how to use it and what the best use cases are. Now, this video could have easily been a multi -hour course walking you through end to end.
But just to make your life easier, in case you're feeling inspired or you're curious about the entire infrastructure and system, I'm going to open source everything that I use to build this little personal assistant. So whether you want to build it yourself or pick and choose and see, what you might wanna pick up next as a personal project, then you can do so as you desire.
You'll find that and a few other guides you can feed your AI agent down in the second link in the description below. And I know this is an atypical type of video for me, but if you wanna see more hands -on experiments with real hardware like this, then let me know by dropping me a like and a comment down below. If you found it helpful, I'd super appreciate it.
I'll see you in the next one.
The Hook
The bait, then the rug-pull.
Most new-model demos give you a game, a website and an app you'll forget by tomorrow. Mark Kashef opens by talking to a $99 desk robot that Claude rebuilt from the firmware up.
Named ideas worth stealing.
04:00list
The eight-part hardware prompt
- Research the exact device, docs and forums
- Break down every component
- Back up the software and firmware
- Goal engineering: dream outcome and services
- Wake word and listening behavior
- Response length and style
- Restrictions and safety rules
- Troubleshooting and tracing
A structure for handing any hardware project to an AI agent, from research through backup, goals, behavior, limits and debugging.
Steal forany hardware or device project you want an agent to take over
How they asked for the click.
VERBAL ASK
11:30product
“I'm going to open source everything that I use to build this little personal assistant... You'll find that and a few other guides in the second link in the description.”
Free Gumroad kit offered as the payoff, followed by a soft like-and-comment ask for more hardware videos.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
