Jev + Claude Code — When you give Voice Control of My Browser to an AI

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Jev + Claude Code — When you give Voice Control of My Browser to an AI · Channel: Prompt Engineer 48 · 18:14 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Claude Code, Prompt and context engineering, AI models

Summary

The video introduces Jev (also referred to in the transcript as Jeff), a system-one model designed to return typed decisions rather than free-form text. The presenter explains its three output types—yes/no probabilities, choices from a defined set, and scores—and demonstrates them on customer-support messages. They show how to access the service, install its Claude Code skill, and use it to build a voice-controlled Chrome browser app. The video concludes with browser-control demonstrations and examples of other tasks involving selection and filtering, while noting that the model is intended for decision-making rather than general-purpose writing.

Key points

The presenter describes Jev as a fast decision model that returns structured, typed outputs instead of natural-language responses.
The video contrasts reinforcement learning from human feedback with what it calls reinforcement learning from calibrated decisions.
The three demonstrated output types are yes/no probabilities (null), selection from a defined set (choice), and a value along a scale (score).
Examples classify urgency, identify the type of Stripe-related problem, and rate its severity.
The presenter claims the model is much faster and cheaper than conventional LLMs, and emphasizes that its output tokens are comparatively inexpensive.
The video explains that the model is not intended for open-ended tasks such as writing poems or explaining topics.
The presenter shows the service’s playground, API keys, credits, documentation, and a skill file for Claude Code.
Using the skill and an API key, Claude Code creates and tests a voice-controlled Chrome app that processes spoken commands and performs browser actions.
Other examples shown include selecting a dress in real time and filtering news items.

Chapters

Start	Chapter	Summary
00:00	Introducing Jev and its decision-focused approach	The presenter introduces Jev as a model that returns decisions in predefined formats and compares its speed, cost, and purpose with conventional LLMs.
03:13	Accessing the service and its three output types	The video tours the service’s access, credits, and API options, then introduces choice, score, and yes/no probability outputs.
04:49	Testing structured decisions in the playground	Through customer-support examples, the presenter demonstrates urgency detection, issue classification, severity scoring, and confidence values.
11:12	Model limitations and Claude Code setup	The presenter reiterates that Jev is a decision-maker rather than a general-purpose LLM, then shows the skill file and its installation for Claude Code.
13:55	Building and running a voice-controlled browser	With an API key configured, Claude Code builds a voice-control app, and the presenter runs it to navigate Chrome using spoken commands.
15:32	Browser-control demo and other use cases	The presenter demonstrates voice-driven YouTube and Wikipedia navigation, then shows examples of dress selection and news filtering.
17:42	Where to try it and closing	The presenter suggests trying the service or accessing it through OpenRouter, points viewers to the available cookbooks, and closes the video.

Quotes

we’re working with decisions, not string. — 00:00

A choice is a system one question type for selecting one option from a defined set. — 04:49

This is not a general purpose LLM. — 11:12

Go to YouTube. Go to Prompt Engineer 48 YouTube channel. — 15:32

any task which needs segregation or ratings or diversion to different branches you can use this structured output — 17:42

Concepts and entities

Jev, TypeSafe AI, Claude Code, skill, Playwright, Faster Whisper

Full transcript (original language)

00:00 Go to YouTube. Go to Prompt Engineer 48 YouTube channel. Click on the first one, Prompt Engineer 48. Go to the videos tab. Scroll up. All right, that’s a wonderful demo. And now let’s go ahead and see and understand what Jeff is. So essentially, we’re working with decisions, not string. And it’s an AI model that never writes a sentence. So you cannot expect to get the answer of write me a small poem or write me a short story or anything like that. It is

00:34 going to respond you in three formats only which we’ll see in just a second. But it is 193 times faster, 244 times cheaper and the output is a typed output which means it responds from the list of things that you have allowed. And this is the summary. 193 times faster than a Frontier LLM to 444 times cheaper on the same work and about $42 per billion input token. Mind it per billion. It’s not per million. And essentially output tokens are free and 150 milliseconds

01:07 decision faster than the human perception which is really great. And we have been using text models as decision engines which is really wrong. So what we were doing uh with all the LLMs uh for example reinforcement learning from human feedback those kind of LLMs the LLMs that we know we put in the prompt and we beg the answer we beg the model in English to answer in JSON format and then we go ahead and parse the markdown French retry when the apologize instead and then we pray that no calibrated confidence so we cannot tell if the

01:41 answer is sure or are we confident about the answer and then we need to babysit uh with a human in the loop and that’s a very bad thing and the automation never actually ships and therefore uh we are not able to automate with the normal LLM. Now what they have done and if we go and now what do Alma has done he has worked with RLHF earlier as well and he spent the last two decades in stealth building this new ways of training models it’s 20 to 200 times faster

02:11 compared to the normal LLM 40 to 400 times cheaper and it the output token is basically free because there is nothing to write in a longer form the answer is just one of the things that you have asked it to answer from and it’s really amazing questions [music] just like you would an LM get the answer. So we have been till now training uh reinforcement learning from human feedback but this new model we are we have not approached AGI and what do you think that is and

02:41 it’s also difficult to automation and it’s also difficult to automate you can see on the left the type save versus any other LLM and what they’ve done for this model is reinforcement learning from calibrated decision and that’s really different you can see the output here it’s really fast it completed in 114 seconds while the normal LM it takes a bit of a time. So you can see they completed in 8.56 seconds. So this is the official website typesafe.ai. This is the first public system one model. System one model meaning it

03:13 thinks fast and the system two models is thinking slow but in the long form. Okay. So you want access to this type.ai. You can go to join weight list here. And I’ve already applied to weight list. I got this $5 of credit. If you go to the uh usage here, we can see that we are here. After you get access, you’ll be brought here. We have home playground usage and you can see that I’ve been using this a lot. We have the API keys here. So, you can generate an API key. Use that API key uh for your creation. So, I can copy this API key and that’s

03:45 done. I’m going to delete this because the API key is visible now. So, if you go to home and if you go to settings here, then you can see that in the billing section, you’ll get about $5 to test it out. And I’ve been testing this a lot and you can see I’ve just spent uh 0.71 and I have a lot to go. I can test so much with this $5 of credit. So, if you don’t have access to this, you can go to open router as well and get access from there. Now, let’s go back to the PPT. So, we took the opposite research direction first. Uh we know that we have

04:18 the reinforcement learning from human feedback. Every LLM you know it’s trained on human perception. It’s optimized to produce words for people. But this Jeff Jev is reinforcement learning CD which is calibrated decisions. It’s optimized to produce typed decisions for machine. Now success looks like a probability is actually right. System one how we know fast judgment not like a slow pros. So this is a system one model. LLM produce words for people but Jeff produce type decisions and it’s really amazing. So

04:49 there are three primitives there are three types of outputs that you get from this LLM. The first is a choice. So the choice is something which will select a particular one from a given list. If you go if you go to the documentation from this button if you go to the documentation and go to the choice you can see that we have these examples. So a choice is a system one question type for selecting one option from a defined set. So what programming languages is this code written. So these are the options and it is going to answer only

05:20 within that format. Okay. The other one is the score here. So this is a degree along a dimension. So this is a probability weighted position. If you go back to score you can see that it gives out this. So for example the uh score question bug severity how severe is the reported bug. So when it says zero it means no impact of functionality one is like broken but two is like blocking issues. So this is a score. It goes from zero one and two. And the last one is a

05:54 null. Null is a yes no question. So it gives a probability of a yes or a no. So if yes has a probability of 90%. No has a probability of 10%. So let’s go ahead and go to this example here. If you go to the playground here and I’ve been playing uh with this on one example, let me show you. So this one you can see that uh we have this statement for example. Hi my stripe is failing. Please help ASAP. I’m losing numerous high value clients. So looking at the

06:24 statement I want to ask three questions. You can send all the three questions together. So the first question that I’m really interested is the is the type of null which means it says yes or no yes or no. So this is u an object here and we we are saying that this is a null type and the instruction the question is does this message express urgency and if I send this you can see on the right yes this is urgent and you can see that this is 90% true and 2% false this is a no

06:56 format okay which is like a yes no or true false so if uh does this sound urgent it’s yes so if I maybe my stripe is please help If I remove this maybe if I remove this then what do you think will happen? My stripe is failing. So it’s like a 60% urgent. Okay. So this changes and it’s not like uh I pre-edited this. This is really fast. So when I send this command it’s really fast. If you go ahead and see the output here you can see the evaluation time in millisecond is 161 millisecond 104 millconds. Really fast. If you send the

07:30 request it’s 118 milliseconds. So it’s really fast. So this is the null type of uh data that you can ask or question that you can ask and the another type is this one. This is the type of choice. So you have to give the choice to this LLM to answer from. So for example the instruction is which option best fit the state the state that the stripe is failing and uh we can put this criteria so accounts billing and payments. So and a small description of the three things

08:00 here. So looking at this question here and given the statement, what do you think the output would be? So if I run this and you can see on the right problem type is payment. Okay, I’m saying my stripe is failing. It looks like a payment issue. If I say my stripe billing is failing and send this, then you can see that the billing is 100% here. It’s like confident 99% confident that it’s the issue of the billing, not of accounts, not of payments. But if I make it like a vague thing, then it looks like account. Okay. Uh and if I

08:34 say uh my customer uh my customer is not able to pay via the via my stripe link. Now what do you think? I think that would be related to payments. So you can see 94% payment related to payment. So this is another question type which is a choice type. You give a choice. You give three choices here and it is going to take it’s not limited to three of course. So it will give out the confidence level as well give out the probability of each of what it thinks.

09:05 Uh and you can see this output here. You can see this output in a JSON format as well. So this is the model uh Jeff 113 and uh this are the answers. So the first answer is this. So no is 0.94. Then we have the problem type of choice and you can see the confidence is 91%age. Then we have payment is 94, accounts is zero. This is the second one. And the third that we’re going to look at is this one. So let’s go ahead and look at the third type of variable

09:35 that you can use. So the third type is the score type. So score goes from 0 to one. So if we put the instruction of where does the severity falls on the scale, then it’s uh I’m losing numerous high value clients. It’s definitely uh on the high scale here. So it’s giving you 1.99. And if I remove this part for example and then send this and you can see that 1.53 is the severity. So this is just an example to show you that you can send all three of these questions together and based on this particular statement based on these type of

10:08 questions that you have asked uh whether it’s a null or let’s say choice and it’s a score you get this output and you can look at the output there is no output tokens as such it’s just that it returns from the things that you have asked whether it’s a yes no which is return in percentage whether it’s a choice it returns the probabilities and confidence and there’s a severity It returns whether 0, 1 or two. And it also returns the confidence level on the choice and the score variable type. Okay. And you can see the input token is

10:40 392 and output token is 70. It’s not zero per se because this output needs uh some tokens to give you. But compared to the other LLM, it it’s essentially free. The output token is free. And there’s the request ID. And this is the evaluation time. You can see 120 mconds. Keeps on changing depending on the servers. 87 milliseconds lesser and it’s like 130. So it’s within like 80 to 150 millconds for this particular type of questions really amazing. So moving back to the website this is the first system one model. So system one model is a fast

11:12 thinking model. Jeff gives AI the properties of code. Now this is really important because uh you can see the speed here. It’s really fast like 193 times faster. Now this is not a general purpose LLM. cannot ask questions uh and ask it to poem, ask it to explain something, but it’s generally a decision maker for you if you already have the options listed out. It’s really important for the software that our AI system use. So, we had the RLHF we have already seen. Now, this is known as reinforcement learning from calibrated

11:42 decisions and it’s 193 times faster, 440 times cheaper. You can see that type safe AI is so fast, it’s really slow and I’ve tested uh the examples here. It’s really it’s really fast and we can see that workflow intelligence versus cost. So average cost is very less. But again you need to remember that the limitation of these type of models. It’s basically a decision maker for you. It doesn’t work like a normal LLM like for example GPT 5.2 Luna would work. So you can see the structured output error rate is very less almost zero error. So there is no

12:15 hallucinations because the output type is fixed. There is nothing to hallucinate. So it’s zero hallucination which is really good. And you can see the price $40 per billion input token compared and you can see 238 times lower input price than Claude Fable 5.1. Really amazing. So let’s go ahead and test this out. Now if you have access to this, you can go to uh playground here uh sorry you can go to home here and you can see that uh we have this option of copying this setup uh to cloud code or aentic options like that. We have the

12:47 skill MD file as well. So you can go ahead and download this uh view here and you can see that uh this is the skill file or I can go ahead and maybe download the skill as a whole. So skill.md I can download this and you can see I’m going to show you this skill file here. So a skill is basically uh the name is type safe AI and there’s a description built with type safe. We have all the documentations listed here and uh it’s basically has everything that you need. So cookbooks and everything is available here. So it’s just a matter of copying this system prompt and putting it on your system. So

13:21 I have cloud code uh here and I what I’ve done is I’ve just pasted this type skill uh type safe skill and we can see that it’s already installed because uh previously what I’ve done is I’ve gone ahead and installed this. So if you go to the documentation here and if you go to the agent skill here for the cloud code you can run these two commands. So if you go to your cmd or command prompt on a Windows system and run this, you can see that it’s already installed here. Okay. So you can go ahead and install this. And once you’re done, this is installed uh into cloud code. So that’s what uh it’s saying that it’s

13:55 already has. Now uh I want to do some examples here. So I ask you to do give some creative examples. So AI as if uh as an if statement. So this is an AI which is like an if statement only. So these are different examples that is given but I really wanted to make that voice control app. So what I’ve done is I’ve given the uh environment file somewhere here. So I put the environment file uh the dot on on the inside thevb file. I put the API key which I obtained

14:26 from the uh location here. And once I have the API keys what I asked it to make that voice control app. It went ahead and made the entire thing for me. uh and it even tested it out on its own. Now what I can do is I can just go ahead and start this up. So I’ve opened that on my code editor uh visual studio code editor and what I can do is that this is the particular folder which it has made this voice chrome here. I go ahead and navigate to the folder then I do pip install-r requirements.txt.

14:58 And once you have uh those requirements done then what you can do is you can just go to the readme file here open preview and you can see this from the mic it goes to the faster whisper it is uh made into text to speech and goes to Jeff direction class then playright chrome local in the CPU on the GPU then we have type validated okay so let’s go ahead and uh start this up so I can say python-m voice chrome and I just run it here this is going to start up my Chrome and I can do voice command now. So, I’ll just stop

15:32 this and start this up again. Go to YouTube. Go to Prompt Engineer 48 YouTube channel. Click on the first one. Prompt engineer 48. Go to the videos tab. Scroll up. Scroll up. Go to the videos tab near the home. Click on the videos tab near the home. Click on videos. >> How much can someone learn realistically in 3 hours all about AI? in just scroll up, click on videos, click on popular, scroll down, down, scroll down. Uh, click on cloud code plus Nvidia Nim. Okay. So, okay. Wait, let me see. Let me

16:07 show you this. So, you can see this is cloud code. Scroll down. Like the video. Click the like button. Scroll down. Down on a new tab. Open Wikipedia. Go to Wikipedia. Wikipedia. Please click on the second link. Scroll down. Scroll down. Go to Pepper Petrella. Go back. Go to pepper petrella. Go back. Click on pepper petrella. Click on pepper petrella. Scroll up. Scroll up. So you can see that this is u a demo which cloud code made for me and this entire

16:38 thing. It’s really amazing to test this out. And you have this little sections of on the right bottom. You can see this one uh when what you said and what are the options that are available. You can see all these and based on these options the action is initiated. And this is one example of what this Jev can do. There are many more examples. Got tons of different examples that people have done. Now, one such example is this uh example of >> hey, I’m going dress try on here. >> How about white?

17:09 >> So, you can see there’s a real time uh dress selection here. That’s amazing. Uh this is another one uh in which you can see that it goes through the news items and it filters for different uh companies or different companies news and it’s so fast. You can see on the right you can see OPUS cloud 5 it’s really slow and the processing speed is slow. So any task which needs segregation or ratings or diversion to different branches you can use this

17:42 structured output Jeff you can go ahead and try it out today go to types safe AI get that $5 of credits or if you don’t get that you can go ahead obviously to open router and use this so install this skill and get started and you can see all the cookbooks um it’s available here you can go ahead and Check this out. And now I will make much more and many more videos on this. I’ve just started creating a few different use cases that would really um indulge you and uh

18:14 wonder you. So until then follow this channel and I will be back with more interesting use cases on Jeff. And in the meantime, you can check out these other videos on my channel. Like, share and subscribe and I will see you next