Big Idea

The argument in one line.

Claude Code's UserPromptSubmit hook can route a prompt to a non-generative decision model first, so questions a database can already answer never spend a single LLM token.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You run Claude Code, Cursor, or Codex against a live database or app and ask it the same kind of factual questions over and over.
You want to cut token spend on requests that are really just data lookups, not reasoning tasks.
You're comfortable pasting a setup prompt and an API key into your coding agent and letting it wire a hook for you.

SKIP IF…

You don't have a codebase or database Jev would have anything to query.
You're looking for a deep explanation of how UserPromptSubmit hooks work internally, not just how to install one.

TL;DR

The full version, fast.

Jev AI, built by TypeSafe, is pitched as a "System One" decision model: instead of generating text token by token, it returns a typed, probabilistic decision in milliseconds. The video wires it into Claude Code through the UserPromptSubmit hook, a mechanism that lets a message get intercepted and answered before it ever reaches the LLM. Every prompt now hits Jev first; if Jev can resolve it with a direct tool call, it responds instantly in a different chat color and no LLM tokens are spent, and only prompts Jev can't handle fall through to Claude. Setup is get a TypeSafe API key, paste their agent setup prompt and skill file into your coding agent, and tell the agent to wire the hook. The demo shows it answering follower-growth and comment-count questions against a live social dataset in about half a second, then repeats it against an uploaded steak-eating survey to prove it works on any dataset, not just the creator's own app.

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 →

Chapters

Where the time goes.

00:00 – 01:21

01 · Intro + demo

Cold open showing Jev answering live follower-growth and 30-day comment questions in seconds, then the creator states the goal: wire Jev into Claude Code, Cursor, or Codex.

01:21 – 02:31

02 · How the hook works

Explains the UserPromptSubmit hook: every message goes to Jev first, Jev answers directly when it can (shown in a different color), and only unresolved prompts pass through to Claude.

02:31 – 06:08

03 · Step-by-step setup

Get a TypeSafe API key, grab the agent setup prompt and skill file, paste both into a fresh Claude Code session, let the agent wire the hook, then test it against a live dataset and an uploaded CSV before recommending a companion UI over Jev's tool calls.

Atomic Insights

Lines worth screenshotting.

Jev AI is described as a "System One" decision model: it returns typed, probabilistic decisions instead of generating text token by token, so it can answer in milliseconds.
The integration point is Claude Code's UserPromptSubmit hook, which lets a message be read and answered before the LLM ever sees it.
When Jev resolves a prompt on its own, the response comes back in a different chat color so you can see at a glance that no LLM tokens were spent.
In the demo, a question about 30 days of social comments returned 3,426 comments, with 2,900 of them already flagged as having received a DM, in about two seconds.
Jev can be pointed at any dataset, not just a live app: the creator uploads a random steak-eater survey mid-video and gets a correct cross-tab answer back in about half a second.
The whole setup is three inputs: a TypeSafe API key, an agent setup prompt, and a skill file, all pasted into the coding agent and wired by the agent itself.
The creator's own recommendation is to have the agent also build a small UI over Jev's tool calls, because seeing the available tools, sort orders, and platforms made the system easier to reason about than the terminal alone.

Takeaway

Route data questions past the LLM before you pay for one

WHAT TO LEARN

A decision model that intercepts a prompt before the LLM sees it can answer simple factual questions in milliseconds and for a fraction of the cost of a full model call.

Separate lookup questions from reasoning questions: a request like "how many comments in 30 days" doesn't need an LLM, it needs a typed tool call against a database.
A hook that runs before the model sees the prompt is the right place to make that split, not a prompt instruction telling the model to "be efficient."
Visually distinguishing which answers came from the cheap path versus the LLM path makes it easy to verify the savings are real and not just assumed.
Testing a new integration against an unrelated dataset (the steak survey) before trusting it on your real data is a fast, low-risk way to confirm it generalizes.
A thin UI over an automated system's available tools and options can make the system's actual capabilities easier to reason about than reading its terminal output.

Glossary

Terms worth knowing.

Jev AI: A decision-model product from TypeSafe that plugs into a coding agent and answers data questions with fast typed tool calls instead of generating a text response.
System One decision model: A model built to return a quick, typed, probabilistic decision rather than generate open-ended text, contrasted here with a normal token-by-token LLM.
UserPromptSubmit hook: A Claude Code hook that intercepts a user's message before the LLM processes it, allowing another tool to inspect or answer the message first.
TypeSafe: The company behind Jev AI, referenced here as the source of the API key, agent setup prompt, and skill file used to wire the integration.

Resources

Things they pointed at.

02:31toolJev AI (TypeSafe) ↗

00:00toolCreator OS ↗

00:00channelCreator University (Skool) ↗

00:00channelKev's No Code Academy ↗

Quotables

Lines you could clip.

00:20

“Boom gets it back bang. We have 89,000 followers. We gained 2127 in the past day.”

concrete, fast payoff right after the setup→ TikTok hook↗ Tweet quote

02:23

“It'll respond in a different color so that we know that it's the Jev response and that we're saving those tokens.”

names the exact mechanic viewers want explained→ IG reel cold open↗ Tweet quote

04:51

“And 36.4% of six-figure earners in this survey eat medium-rare steak.”

funny, specific proof-of-concept payoff→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

So last night I connected this new AI tool called Jev to the cloud code workspace that runs my entire business. And now this Jev AI tool can answer any question about my code base or my database in literally a matter of seconds, no matter how much data it has to process. Check this out.

Tell me about my follower growth for the past week. Boom gets it back bang. We have 89 ,000 followers.

We gained 2127 in the past day. Well, that was what literally like three seconds. Let's try another one Can you tell me all the comments that we received on all of the socials for the past 30 days?

All right, that took like two seconds, 3 ,426 comments in the last 30 days, 2 ,900 of them got a DM comment sent to them from one of my videos. And here we can see all of the usernames. Like this is actually ridiculous guys.

And that is just the beginning of this Jev rabbit hole. But before I unleash an absolute swarm of Jev AI tutorials for you guys. First thing we have to get out the way is how to set up Jeff with Claude code.

So the rest of the tutorials are just going to be easy sailing. And so that's what the purpose of this video is today. I'm going to show you step -by -step how to get the Jeff AI connected to Claude code or cursor or codex, whichever model that you use in literally a matter of like five minutes.

So let's dive right in. So I'm going to show you guys today is how to get Jeff actually integrated inside of your coding agents terminal so that every single message that we send in actually goes to jev first and this is all possible to this thing called the user prompt submit hook this is the only complicated part of the entire video but essentially the way that this works is that a message gets sent in and it goes through this hook first and this allows jev to read the message before claude or codex reads the message and if it decides that it can get the task done without using the ai model it'll get it done and the way that i've designed it while using it is that when Jev responds back to us in the chat, it'll respond in a different color so that we know that it's the Jev response and that we're saving those tokens.

And so let me just show you that example quick and then we'll start building. So if I send in a message that says, tell me the social account follower growth in the past 30 days, that's going to hit Jev and Jev's going to realize that it can actually just make a quick tool call. And that's why we're getting that message back in yellow, right?

And so that is a completely Jev responded message in the Claude CLI chat. and that's what we're going to be doing today which is absolutely baller everything that you need is going to be uh in the description down below but essentially the actual tools that you need is your jev ai api key and they have a waitlist form but the waitlist form literally you get approved in like 10 minutes and it's so freaking affordable that like if you load it with one dollar you're gonna be good for the entire month from there you just want to choose whatever coding environment you're comfortable with then we can get started so once you have your api key you want to save that somewhere and then you want to head over to the website uh for the of the company behind jev where they give you a full agent setup prompt and you just want to copy this agent setup prompt as well as the skill md and that alongside with your api key you can now submit to your coding agent and so now in real time i'm going to show you exactly what that looks like so i'm going to head over to cursor i'm going to click new window i'm going to open new folder we're going to call this jev ai plus claude

we're going to click create all right so now we've got the cloud terminal open i'm going to snap my fingers with all of that information that i told you to add into the chat now essentially we just wait for the agent to get the skill loaded get the api key initialized and from there you're basically set up and so now the fun part can begin because all we have to do now is instruct claude to set up the hook so that every prompt goes through jev first and i have the prompt already written out for you guys so you can just grab it at the link down below and i'm just going to literally just drag it and drop it into here and we're going to let the agent cook it up and that should literally turn this chat that we're working on right now into a jev first chat and then we'll test it out with a prompt so after the agent wired in that setup prompt so that the prompts now go through that hook and goes to jev first i told the agent to now change all of the messages that we get back from Jeff to green so that we can tell that Jeff's actually activated.

So now if you followed up up until this point in the tutorial, we have Jeff integrated into our cloud agent in a repository that has no data. And so there's nothing that the Jeff agent can actually do. So what I'm going to give you guys is this random survey of steak eaters.

And it's just random information about people that eat steak and how well they like to have it done and how it correlates to their life. It's very random. Awesome.

So now that Jeff can actually see this data set, we can now start asking questions. And so the first question I'm going to ask is what percentage of six -figure earners eat medium rare steak? I'm going to send that in.

And in about half a second, we got it back. And 36 .4 % of six -figure earners in this survey eat medium -rare steak. how ridiculous and so that guys is how you set up jev with cloud code but what i recommend is also then telling your agents to build a ui or user interface so you can actually see all of this logic that jev has behind the scenes and it actually helps you understand it a bit easier and if i were to give you an example of what that looks like if we head over to my operating system and we go over to the jev section we're going to be able to see that at the top each choice each set of choices is slowly starting to have more options so like it has 11 tool calls it has 22 different dates and times that it can choose from it can sort information in six different ways there's eight social platforms and there's four statuses and it'll just continue to add things to this and so that's why like this jeff for example can answer so many different questions i could be like what comments did we get today on instagram specifically and it'll literally like bang out this answer in literally half a second with

it costing any tokens and so i'm really happy that i was able to get this tutorial out because now i can start showing you guys all of the insane workflows that this has unlocked catch you guys in the next one

The Hook

The bait, then the rug-pull.

Last night the creator wired a new decision-model tool called Jev into the Claude Code workspace running his business, and it started answering follower-growth and comment questions in about three seconds flat, at zero LLM cost. This video is the wiring instructions.

Frameworks

Named ideas worth stealing.

01:31concept

Prompt-first interception via UserPromptSubmit

Every prompt sent to the coding agent is routed to Jev before the LLM sees it. If Jev's decision model can resolve the request with a direct tool call, it answers immediately and the LLM is never invoked; otherwise the prompt falls through to Claude as normal.

Steal forAny coding-agent workflow with repetitive factual/data lookups that don't need reasoning

CTA Breakdown

How they asked for the click.

VERBAL ASK

00:00link

“all my links... Get Creator OS and the API here, Work with me 1 on 1, join Creator University”

Standard description-block plug for the creator's own Creator OS product and coaching/community links; no in-video verbal pitch beyond pointing to "the link down below" for the Jev setup prompt and data.

MENTIONED ON CAMERA

02:31toolJev AI (TypeSafe) ↗

00:00toolCreator OS ↗

00:00channelCreator University (Skool) ↗

00:00channelKev's No Code Academy ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.