How to Use Jev AI Better than 99% of People

0:00Jeff just crushed Claude and Open AI by becoming one of the fastest [music] and cheapest model out there. But, that's not all of it. Jeff is engineered to bring a totally new approach to working with AI. And [music] in this video, I'll show you exactly how to use and master it by walking you through some of the most useful examples I could find. And all of them are jobs where the AI you use every day wastes a ton of your time and money. Pretty much every artificial intelligent model out there right now, like ChatGPT, Claude, or anything else, is a prediction machine at the most basic level. You give it a

0:39message, and based on patterns it learned from a huge amount of text, it predicts the best reply one word at a time until it's done. And honestly, it's an incredible system because that's the reason most of these models can work. But, these AIs are designed to process massive amounts of information all the time. Even when you only need them to make a small decision, like sorting through all of your emails and figuring out which ones are worth replying to and which ones are urgent, which ends up costing you time, resources, and money just to get one answer. And a lot of the time, like with those emails, all you

1:17actually need is a simple yes or no. That's exactly the problem Jeff was created to solve. It doesn't even know how to write long answers or code like traditional LLMs do. Instead of predicting words, it's built to make choices. You give it something to look at and a short list of possible answers, and it works out how likely each one is to be right, all at the same time. So, you don't get a reply at all, just the answer that fits best and how sure it is about it. For example, if you get thousands of emails coming to your inbox every single day, Jeff can go through

1:49all of them and pick out the ones that actually need you based on the rules you set. And that's where it beats a normal AI because for each email, it only chose one answer instead of writing out a whole response or thinking too much about it. So, it's much faster and costs [music] far less. Now, there's a catch. At the time of recording this video, Type Safe, the company behind Jeff, has paused sign-ups, which means you can't

2:12get in through their [music] websites. But, there's another way in through an API from a company called Vercel. [music] Vercel lets you use lots of different AI models from one account, and Jeff is one of them. I'll connect my Vercel account to Cloud Code, which lets it reach Jeff from there. And all the prompts I'm using in this video are free in the description, [music] including one you can paste into Claude, ChatGPT, or any other AI that walks you through this whole setup step-by-step. Also, if you want to go further with AI than this, my school community, AI Fluency, takes you from casually using ChatGPT to using AI

2:48like the top 10% of people in 12 weeks. You get a new video every day for 3 months, over 1,000 done-for-you resources, and help from me and my team every day inside the community. Just use the link down in the description to join us. So, whenever you need AI to make a quick pick from a few options, that's the job Jeff was built for. But, if Jeff is the one doing the work in this video, you're probably wondering why we need Cloud Code at all. It's because it can't make software by itself. So, I'm using Cloud Code to build the apps I'll use, and Jeff to do the actual work inside

3:21them. And Cloud Code handles connecting Jeff to the app, setting up the rules, and fixing anything that breaks. So, I don't have to do any of it by hand. If you're thinking about using Jeff, you probably have Claude or ChatGPT on your computer anyway, so it makes sense to get the best of both worlds. Now, I'll open Cloud Code inside the folder that's connected to Jeff and paste in this prompt, which builds a supporting box for a made-up coffee brand called Kettlewick with 200 customer support emails to sort. Once it's built, I'll open it up, and right now it's completely empty. All I have to do is

3:54click start in the top right corner and Jeff immediately goes through all 200 support tickets one at a time sorting them into these columns. And by the end it's all organized and I can see exactly which problems people are having and which ones to fix first. For example, if I open any card it tells me which team Jeff picked and how sure it was about every option. It also shows how urgent the email is, whether the customer wants a refund and how long Jeff took to

4:20decide, which is usually under a second. The way this works is that for every email the app sends Jeff the subject, the message, and the customer's plan as separate labeled parts so it can tell which part is which. Then it asks three questions at once, which team should handle the problem, how urgent it is, and whether the person is asking for their money back. The team question also has an other option, so an email that

4:44fits nowhere still has somewhere to go. After that a few simple rules decide where each email ends up. If Jeff is at least 60% sure about the team, the email goes straight into that team's column. If it isn't sure or the email doesn't fit any team, it goes to human review so a person can check it. And if a customer wants a refund and they're blocked or angry, the card gets a red refund risk badge so it never gets missed. All 200 came to about half a cent at Jeff's normal price. Running the same 200 emails through Claude Haiku, one of the cheapest chatbot models, would come to

5:17roughly 20 cents, which is about 40 times more. And since [music] an inbox like this runs day after day, that gap grows with every support ticket that comes in. To use this for real, you can have Claude Code connected to whatever help desk you already use for your software or service or just swap in your own business details. Then every single day you get a sorted list like this that's easy to skim so you can deal with the important ones first. So that's 200 support emails sorted in a few minutes and I never had to read a single one myself, unless it ended up in human

5:49review. And the next use case is for anyone who's into AI content creation and has to write a lot of prompts. I'll paste in this prompt into Claude Code, and it builds a prompt checker for AI video and image prompts in one go, connecting it to Jeff for me. If I open it, there's a box for my prompt [music] and a few examples I can load underneath. I'll click the one with the woman in the red raincoat, and a moment later the whole prompt is checked with a bar and a score for how ready it is to generate. [music] And it's super fast because Jeff answers

6:17every check at the same time, so the whole list loads in about a second. Below that, it shows me what's included in my prompt alongside what's missing. And above the score, it tells me what kind of prompt [music] it thinks it is. For example, if I click the one called mountain lake at sunrise, it's a really

6:33short prompt without much detail in it. So, it only gets a score of 27 with three things included and 13 missing. Similarly, a perfume bottle on a table gets an even lower score with just the subject included and 13 other things missing. And that's what makes this so useful when you're writing prompts, because instead of guessing, you can see exactly what your prompt still needs and

6:55keep adding to it until it's complete. The way this whole thing works is that whenever I stop typing for half a second, Dev sends my prompt to Jeff with a list of questions. First, what kind of prompt this is and whether it's for a video or an image. Then, a set of simple yes or no questions like does it mention

7:13a lens or does it describe the lighting? Dev only shows the ones that matter for that kind of prompt, [music] and anything missing comes with a short example of what you could add. Just keep in mind that Jeff reads your own checks literally, so write them exactly the way you mean them. Each check costs a tiny fraction of a cent, which matters here because it runs again every time you pause. I also had Claude Code add

7:37another feature to this, which is batch. So, I switch from check to batch in the top right corner and click load 10 sample prompts. This is made for when you've written a whole set of prompts and want to check all of them at once. Then, I'll click check all, and in a few seconds, it gives me a summary of every prompt with its type, how red it is, and what it's missing, with the weakest ones at the top. This is incredibly useful for someone like me who works in AI content creation and has to write top-tier prompts to get the best results. The next use case goes through

8:08a full podcast episode and picks out the best moments to turn into short videos. To do that, I'll paste this prompt in the prompt code, and it builds an app that scores any YouTube video I give it. In the prompt, I also added the link to this Diary of a CEO episode with Drew Ski, so the app opens with it already loaded. The timeline under the player is the full episode split into small parts, about 30 seconds each. Jeff goes through those parts one by one and gives each a score from cold to hot, which just means how good that part would be to post on

8:40its own. Cold means bad, and hot means good. If I hover over the timeline, I can see what's being said there and how deep on each of the four checks I set for Jeff. And if I click any of the top five, the video starts right at that moment, so I can see for myself if it's any good. In the bottom right corner, I can also change how much each check counts, and the top five re-ranks instantly to match what I care about most. This one is almost 2 hours long, and scoring every second of it cost less than a cent. And whenever there's a new

9:11episode I want to go through, [music] I just paste its link into this box, and Drew handles the rest automatically. If you regularly make short videos from long podcasts like this, this cuts out almost all of the sorting. So now, any podcast I paste in gets its five best moments picked out by my own rules, and I never have to sit through 2 hours of it to find them. And the last use case is for anyone looking for a job, because this reads your resume against every active startup from Y Combinator and shows you which companies you'd fit best. To do that, I'll paste this prompt

9:47into Cloud Code and it built me an app that pulls in every active Y Combinator company, which comes to 4,322 of them. It also wrote a sample resume for a made-up person called Maya who works in AI automation and content. But you can paste your own resume into this box instead. Right now, the grid is empty, so I'll click start and Jeff begins reading straight away. As it goes through each company, it gives that square a color based on how well it fits the resume. Dark means there's no overlap at all and this light green means a really strong fit. Once every company is read, Jeff takes the top 100

10:27it sorted and checks them again much more closely. And those are the squares with a white border. At the end, I get a top 25 short list with a fit score and a link to each company's website. I can also hover over all these little squares, which look kind of beautiful together, and each one shows me which

10:44company this and how well it fits me. Think about that for a second, because that's every active startup in Y Combinator read against one resume in under 10 minutes. Running the same job on one of the usual frontier models would have taken a lot longer and cost a lot more. The first pass is quick because Jeff only asks how well Maya's skills fit what each company does. The second pass asks three separate things

11:10about the top 100. Do her skills match? Would the company actually need someone like her? And does the location work? And with these sliders, I can decide what matters most. So if remote work is a must, I push that one up and the short list reshuffles instantly with some companies climbing to the top and others dropping down. So if you're job hunting or you're a freelancer looking for clients, you can paste in your own resume and walk away with a list of companies worth reaching out to. Just keep in mind that it only knows what's on each company's public YC profile, so it can't tell you which roles are

11:45actually open right now. And that's what you can do once you know which AI to use for which job. This is exactly what I go deeper into inside my school community, AI Fluency. So, click the link down below to join. [music] Thanks for watching and I'll see you in the next one.