Claude x Jev skill
Skill that hands bulk pick-one, yes-or-no and scoring questions to the Jev model while Claude reads only the uncertain items.
- 97 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Text is sent to the third-party OpenRouter service, and the /jev-gate hook makes decisions about tool calls.
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/claude-x-jevDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Text is sent to the third-party OpenRouter service, and the /jev-gate hook makes decisions about tool calls.
This is third-party code. Review the repository files before installing.
What it does
Jev is a decision model from TypeSafe on OpenRouter. You hand it text and a set of choices, and it returns a pick, a probability per choice and a confidence number. It cannot write prose. The skill puts Jev in front of Claude: Jev sorts, checks and scores, Claude reads the disputed remainder and writes the replies. The pack ships 13 commands, including `/jev-classify`, `/jev-check`, `/jev-score`, `/jev-route`, `/jev-verify` and `/jev-gate`. The last one installs as a PreToolUse hook and approves, blocks or escalates a tool call. The commands also run outside a Claude session, from cron, n8n or a script.
Who it is for. People who triage mail, comments or leads in large batches and do not want to spend a Claude request on each item.
Good fit when
- Incoming mail has to be sorted into fixed labels with confidence for each one.
- A Claude draft is checked against the thread and your rules before sending.
- Agent tool calls pass through a quick safety filter.
- Two company records are checked for duplicates.
Not a fit when
- You need prose, a reply or a summary, since Jev returns only labels and numbers.
- The question is open and no label set is defined.
- There is a single item that Claude is already reading.
Example request
Sort these 200 emails into the email-triage preset labels and show only those where Jev confidence is below the threshold.Limitations
You need python3 version 3.8 or newer and an OpenRouter key, and five dollars of credit is enough. The text of emails, comments and leads goes to OpenRouter and the Jev model, so send private data deliberately. Jev answers each question on its own and does not link answers, so contradictions are resolved by a rules block in the preset. The comparison figures against Claude Opus come from 216 emails in the author inbox, and other data will differ. The confidence threshold is picked by `/jev-tune` from examples you labeled.
How to disable. Delete the skill from the Claude Code skills directory and remove the PreToolUse hook if you installed `/jev-gate`. The OpenRouter key stays with you, so revoke it in the service dashboard.
FAQ
How does Jev differ from a normal language model?
It does not write prose. It takes text and a list of choices and returns the pick, probabilities and a confidence number. The same input always gives the same answer.
How much does it cost?
Per the README, input costs about 4 cents per million tokens and output is free. Five dollars of credit cover hundreds of thousands of decisions.
What does /jev-gate do?
It installs a quick check in front of tool calls as a PreToolUse hook. A decision takes about 0.3 seconds: allow, block or escalate to a person.
How do I choose the confidence threshold?
The `/jev-tune` command picks a threshold from items a person has checked. Confident answers are handled automatically, the rest go to Claude.
Does it need a Claude session?
No. The commands run from cron, n8n, a hook or a plain script, and `/jev-watch` runs a preset on a schedule.
Discussion
No questions or reviews yet.
Related
SkillsA Claude Code command set that finds jobs, fills application forms in Chrome and sends only the ones you approve.
A plugin that makes Claude Code a background assistant: scheduled tasks, Telegram, Discord and Slack replies, a web dashboard.
Full task-lifecycle skills for Yandex Tracker
yt-bp-claud-code-skills
Seven Claude Code skills that carry a Yandex Tracker issue from plan to merge: estimation, work, review, deploy, tests, merge request
A Telegram bot drives Claude Code on your computer: you message the chat and the agent edits the project and streams the reply.
A Claude Code skill that collects BOSS Zhipin vacancies, matches them against a resume and can apply to the best ones.
Orchestrator for Claude Code, Codex and Gemini CLI: YAML workflows that loop until your own quality gates pass
README in short
The skill splits work between Jev and Claude: Jev makes typed decisions in a fraction of a second while Claude keeps reading, writing and judgment. Six presets cover email, comments, leads, duplicates, drafts and a tool gate.
README badge
Original SKILL.md
--- name: claude-x-jev description: Fast, cheap, typed decisions for Claude Code through Jev, TypeSafe's decision model on OpenRouter. Jev sorts, checks, scores, gates and verifies at 0.3 seconds and a fraction of a cent per item, with a confidence number on every answer; Claude keeps the reading, writing and judgment. Use when the user says "jev", "/jev", "/jev-setup", "/jev-classify", "/jev-check", "/jev-score", "/jev-route", "/jev-verify", "/jev-gate", "/jev-match", "/jev-tune", "/jev-lint", "/jev-bench", "/jev-watch", "/jev-status", "classify these", "sort my inbox", "triage these comments", "qualify these leads", "score these", "is this safe to run", "gate tool calls", "verify this draft", "are these the same company", "which threshold", "jev vs claude", "decision model", "openrouter decisions", or wants to run a yes/no, pick-one, or rate-on-a-scale question over many items without spending a Claude session on each one. type: standalone version: 0.1.1 category: operations configured: false allowed-tools: [Read, Write, Edit, Glob, Grep, Bash, WebFetch, AskUserQuestion, Skill, Task] metadata: version: 0.1.1 author: Charles J Dove homepage: https://www.charlieautomates.com/ repo: https://github.com/charlesdove977/claude-x-jev community: https://start.ccstrategic.io/skool --- <activation> ## What A division of labor. Jev is a decision model: hand it choices and it picks one, with a probability on every option and a confidence number on the pick. It never writes a sentence. Claude reads, thinks, and writes. This skill puts Jev in front of Claude as the sorter, so Claude only reads what needs reading. Measured on 216 real inbound emails (September 2026): Jev sorted them in 0.3 seconds each for one cent total. Claude Opus needed 1.3 seconds each and $4.73. They agreed on 87 to 94 percent of labels. When Jev's confidence was 0.7 or higher (82 percent of the inbox) they agreed 95.5 percent of the time. Below that, a coin flip, and Jev flagged every one. Three primitives, every command is one of them plus a routing pattern: | Primitive | Question it answers | Comes back | | --- | --- | --- | | `choice` | Which one of these? | the label, a probability per label, a confidence | | `noul` | Does this hold? | a probability of yes | | `score` | Where on this ordered scale? | a weighted position, a probability per level, a confidence | ## When to Use - Many items, one narrow question each: classify, triage, qualify, rank, dedupe - A gate in front of an action: should this tool call, send, or delete go ahead - A check on Claude's own output before it ships: is this draft grounded, on-rules, on-tone - Anything that has to run outside a Claude session: cron, n8n, a plain script ## Not For - Writing anything. Jev returns labels and numbers, never prose. Drafts stay with Claude - One item where Claude is already reading it. The win is volume and speed - Open-ended judgment with no fixed label set. Define the labels first, then Jev </activation> <persona> ## Role Decision-systems engineer. Turns a fuzzy "sort these" into a typed question set with thresholds, and refuses to trust a threshold that has not been checked against real data.