How to Use Jev With Claude for Faster Workflows

TL;DR

Use Jev for small, repeatable judgment calls that can be returned as probabilities, while Claude handles uncertain cases, planning, and rule refinement. Build a self-improving loop, include a fallback category when appropriate, and choose the browser automation strategy based on task complexity rather than asking Jev to execute one rigid end-to-end plan.

Transcript

Jev has taken the world by storm over these last few weeks. And that's because it is not only 200 times faster than the best that OpenAI or Anthropic have to offer. It's also 400 times cheaper. But Jev is not a chatbot. This isn't something you can just talk to with natural language and get a sentence back. And because of that, most people are usin... Read More

Key Insights

• Jev is a probability-producing model rather than a conversational chatbot. It takes a question and returns confidence scores, allowing workflows to use its output for quick decisions without paying for a full natural-language response from a large language model.
• Jev is presented as 200 times faster and 400 times cheaper than leading OpenAI or Anthropic offerings. The stated price is about $43 per billion tokens, making it attractive for workflows that perform large numbers of small, repetitive judgments.
• The best use case for Jev is a small, repeatable judgment call. Examples include deciding whether a customer message belongs with billing, technical support, or sales, and sorting incoming email into categories such as brand deals, agency pitches, personal messages, or automated mail.
• A self-improving loop is essential because Jev's probabilities depend on the quality of its rules. When a result falls below the selected confidence threshold, a stronger model can resolve the case and recommend a rulebook change for future decisions.
• The email classification workflow improved from handling 43% of messages on its first day to 81% by its tenth day. It used a 90% confidence requirement, escalating lower-confidence messages to Opus while using those cases to refine the classification rules.
• A Jev request can combine the item being evaluated, a specific classification question, and a rulebook defining every category. These definitions provide the task-specific logic behind the returned probabilities and can be revised as uncertain or incorrectly handled examples appear.
• A none-of-the-above option can prevent misleading confidence scores when no provided category fits. This fallback is not necessary in every scenario, but it gives Jev a valid destination for unfamiliar inputs instead of forcing them into an inappropriate category.
• Browser automation can use Opus for short planning cycles and Jev for rapid execution. In the Google Flights example, this loop finished in 12 turns versus 36 turns for the Opus-only control, while a single comprehensive plan handed to Jev became stuck.

Questions & Answers

Q: What is Jev and how does it produce answers?

Jev is described as an AI model that returns answers as probabilities instead of generating full conversational sentences. For example, it can evaluate a customer message about an unwanted subscription charge and assign a 92% confidence score to the billing category. This focused output is what allows it to make repeated decisions quickly and inexpensively.

Q: When should Jev be used in an AI workflow?

Jev should be considered when a workflow contains small, repeatable judgment calls. Suitable examples include routing support tickets, classifying messages, sorting emails, and choosing individual actions during some browser automations. The model is most useful when the required output is a probability or category decision rather than an explanation written in natural language.

Q: Why use Jev instead of a standard large language model?

The main reasons presented are speed and cost. Jev is claimed to be 200 times faster and 400 times cheaper than leading OpenAI or Anthropic offerings, with a stated cost of about $43 per billion tokens. Those differences matter when a workflow repeatedly performs simple classifications that do not require a full sentence from a conversational model.

Q: How can Jev and Claude create a self-improving workflow?

Jev can make classifications when its confidence exceeds a chosen threshold, while uncertain cases are escalated to Opus. After resolving an uncertain case, Opus can inspect Jev's existing rulebook and propose an edit that improves future decisions. Repeating this process turns deployment into an ongoing feedback loop instead of a one-time configuration exercise.

Q: What information should be included in a Jev classification call?

The email example sends three components to Jev: the email being classified, a question asking which category it belongs in, and a rulebook defining the available categories. The listed categories include brand deal, agency pitch, service pitch, personal, automated, and none of these. Clear definitions help make the resulting probabilities useful for that specific workflow.

Q: Why should a Jev workflow include a none-of-the-above option?

A none-of-the-above option gives Jev a legitimate response when none of the defined categories matches the input. Without that outlet, the model may be forced to distribute probability across unsuitable choices, producing confidence that appears stronger than the available options justify. The presenter notes that this fallback is helpful where appropriate, although it does not fit every scenario.

Q: What is shadow mode for introducing Jev?

Shadow mode lets Jev observe how Opus performs a task before Jev begins making operational decisions on its own. In the email example, Opus could sort messages for 10 days while Jev watches the decisions. Opus then uses that history to produce or refine a rulebook, after which Jev can begin handling classifications independently.

Q: How should Jev and Claude divide browser automation work?

The division should depend on the task rather than follow one fixed pattern. In the Google Flights test, having Opus plan the entire sequence before Jev executed it became stuck. A loop worked better: Opus produced a small plan for each decision, and Jev executed the corresponding action. That approach completed the task in 12 turns, compared with 36 for Opus alone.

Summary & Key Takeaways

• Jev is designed to answer questions with probabilities rather than conversational sentences. That limited output makes it suitable for fast, inexpensive decisions such as classifying customer messages or sorting emails. The presented figures claim it can operate 200 times faster and 400 times cheaper than leading OpenAI or Anthropic offerings.

• Effective Jev workflows require an accurate rulebook and an escalation threshold. In the email example, classifications below 90% confidence were sent to Opus, which reviewed the outcome and proposed rulebook improvements. The share handled by Jev rose from 43% on the first day to 81% by the tenth day.

• Browser automation works best when responsibilities match the task. A rigid plan created entirely by Opus and then handed to Jev failed during a Google Flights test. A loop in which Opus created small plans and Jev executed individual actions completed the task in 12 turns instead of the control's 36 turns.