JEV + Claude Code: 7 Marketing Workflows With Setup

Marketing runs on thousands of tiny decisions. Which lead is worth contacting? Which message fits that lead? Which ad creative should we test? Is this search term coming from a buyer? Is this ad starting to fatigue?

Right now, humans judge those one by one, or teams stuff huge amounts of context into an LLM and pay for an essay every time. Both are slow and expensive.

JEV fixes the decision part. Here's how it works, how to set it up inside Claude Code, and seven marketing workflows you can copy: three we run for our business and clients, and four from other builders worth testing.

JEV vs an LLM in One Paragraph

An LLM is built to write. You give it a prompt and it produces an answer one token at a time. JEV, from TypeSafe AI, is built to decide. You give it a situation and the possible outcomes, and it returns a structured answer with probabilities. Ask about a suspicious invoice and an LLM writes a paragraph. JEV returns something like clean 80%, needs review 7%, fraud 5%, and that output can trigger the next step immediately. That's why it's so much faster and cheaper for judgment work.

If you want the full explainer, read What Is Jev AI?. This post is about putting it to work.

The Three Building Blocks

Block	What it does	Marketing example
Choice	Picks one option from a list you give it	Is this email a sales lead, a sponsorship, support, or spam?
Know	Returns the probability that a statement is true	Does this lead show buying intent?
Score	Places something on a scale	Hook strength from 0 to 100

You can ask many questions about the same input at once, in parallel. For marketers, that means scoring hooks, spotting buying intent and qualifying leads without asking an LLM to write an essay every time.

JEV doesn't replace your LLM. They work together: the LLM writes, JEV decides. The result is a much lower cost to run agents.

Set Up JEV Inside Claude Code

This works in Claude Code, and in other agent harnesses like Codex and Hermes.

Add the official TypeSafe AI skills to Claude Code and install them.
Get an API key. You can get one from TypeSafe AI directly, and it's also reachable through OpenRouter.
Add the key to your environment, then ask Claude Code to use JEV for the decision steps in your workflow.

That's the whole setup. The interesting part is deciding which questions to ask.

Workflow 1: Inbox Triage

Most of us get hundreds of emails a day, each with a different priority and urgency. Your inbox doesn't sort by either.

We send each email to JEV once and ask seven questions at the same time:

Category
Urgency
Revenue potential
Whether it needs your reply
Tone
Phishing risk
Next action

What comes back is a ranked inbox you can act on:

A sales lead at 80% confidence: reply today.
"Our agency just quit, we need help": urgent and revenue related, so book a call right away.
A "security team" email saying your account will be suspended, from an unofficial sender: flagged as phishing.
A brand sponsorship: reply today. You can even automate the reply with your media kit and rates.

The next step is to send only the important ones to Slack, Discord or Telegram, so you stop living in your inbox. Our agents run this for us, and because JEV handles the decisions, the cost is tiny.

Workflow 2: Lead Scoring Before Outreach

Lead generation has many steps: sourcing, verifying contact details, enriching, categorizing, and putting people into campaigns. The step that matters most is deciding who deserves your attention.

For each lead, JEV reads the title, source and message, and answers six questions:

Fit
Pain
Buying intent
Two red-flag checks
Next step

The score and ranking are calculated on the page, so you can change the weights without calling JEV again.

For us, fit means we work with founders directly, so a manager title scores lower. Pain means they need video for their business. Buying intent comes from signals like fundraising, recent news, hiring, or frustrations they shared on social media.

Then we route by priority:

High fit, clear pain, clear intent: top queue, contacted first.
Fits the ICP but no clear pain or intent: lower-priority inboxes.
Selling to us, partnering, or not buying: removed.

This matters most when inbox capacity is limited. We run more than 100 outreach inboxes and it still helped, because every send goes to someone worth the slot.

Workflow 3: YouTube and Short-Form Intelligence

We produce videos for founders, so the intelligence layer is the most important part of our system: which hooks work, which formats work, and how to structure content so a client becomes known in their niche.

Our agents scraped more than 200 YouTube videos and ran more than 3,000 JEV judgments on them to understand why some videos on the same topic die while others break out. Here's what we found for our topic:

Numbers in the title or thumbnail help. Specificity wins.
Strong emotion or no face beats a calm face. I always used a calm face, which is probably why some of my videos underperformed.
Hype words hurt. "Insane" and "wild" work on social feeds but not on YouTube.
Explainer hooks win. We classified the first 30 seconds of each video as explainer, complainer, hype, news or tutorial, and explainers performed best. That's why this video is an explainer.

Then we score our own title and thumbnail ideas against those patterns before publishing: curiosity gap, specific payoff, hype words, clarity, first person, and proof. A neutral face with a vague title scores low, and now we know why before we film.

For Shorts, we use JEV in three places:

Ideation: finding winning concepts, formats, hooks and structures.
Learning loop: after publishing, we run the performance data through JEV to see which hooks, topics and formats worked, then save those patterns back to our agents so they pick better ideas next time.
Production: deciding which generation tool fits each client video.

The scoring layer we used before cost us more than $100 a month. JEV does the same job a lot cheaper.

Four Community Workflows Worth Testing

These come from other builders. We haven't run them ourselves yet, but each one looks legit.

Workflow 4: Live Sales Call Copilot

It listens to the call, tells you what to say next, helps you handle objections, and shows the stage of the call with a live probability of closing. JEV judges fast enough to keep up with the conversation. Read the case study.

Workflow 5: Lead Finding on Reddit

Being the first to respond to an inquiry or complaint on Reddit starts real conversations, but scanning several subreddits with an LLM gets expensive. JEV decides whether each post is relevant to your offer and labels it as a complaint, inquiry, question or general post, so your team knows what to answer first. Read the case study.

Workflow 6: Ad Creative Analysis

Tag every live ad in your category's Meta Ad Library by hook, format and offer, find the patterns, let an LLM write the briefs, then use JEV to score each hook's survival odds and catch fatigue. Read the case study.

Workflow 7: SEO and GEO Research

Much of SEO and GEO work is pulling and processing data from many sources to find gaps. A judgment layer saves time and tokens on every page decision. Read the case study.

Bonus: JEV Tools for Claude Code

A few builders have made general tools on top of JEV:

Skill loading: keeps Skills out of your context window until they're needed. Useful if you have a lot of Skills. If your Skill set is small and clean, you may not need it.
Model routing: picks a model based on the complexity of each task, so repetitive work goes to cheaper models. It doesn't invalidate the cache in your main thread.
Fast compaction: compacts context faster when you're near the limit. See instant compaction.
Browser use: quick browser actions, like pulling data from YouTube for content monitoring. See Browser Use plus JEV.

Where JEV Fits

JEV is new, and I don't usually cover new tools unless they change how businesses run. This one does, because it separates the judgment layer from the generation layer. Let the LLM write. Let JEV decide.

For the full set of case studies and our video scoring prompts, see the JEV guide.

Frequently Asked Questions

What is JEV used for in marketing?

JEV handles the many small decisions in marketing: sorting emails, scoring leads, rating hooks, titles and thumbnails, spotting buying intent, and catching ad fatigue. It returns a choice or score with a probability instead of writing text.

How do I use JEV in Claude Code?

Install the official TypeSafe AI skills in Claude Code, add your API key, and ask Claude Code to use JEV for the decision steps in your workflow. The same approach works in Codex and Hermes.

Where do I get a JEV API key?

From TypeSafe AI's website. It's also available through OpenRouter.

Does JEV replace ChatGPT or Claude?

No. LLMs still write. JEV makes the decisions around that writing, which makes agents faster and cheaper to run.

Is JEV cheaper than using an LLM for scoring?

For judgment work, yes. It returns a structured answer instead of generating an explanation. Our old scoring layer cost more than $100 a month, and JEV does the same job a lot cheaper.

Want this running in your business? Book a call and we'll show you how we set up the decision layer for your marketing.

JEV + Claude Code: 7 Marketing Workflows (Setup + Examples)