<aside> ⚡
Let your agent plan. Give the queue to Jev.
Astra or Claude writes the decision rules, Jev works through the records, and your agent reviews the exceptions. This kit gives you the skill, a Python runner and 100 synthetic emails to try.
</aside>
Noul returns the probability of yes. Use it for spam, buyer intent or “does this need a reply?”
Choice selects from your named options. Use it for the owner, queue or sponsorship category.
Score places the record on your written scale. Use it for urgency or budget tier.
Unzip it and open START-HERE.md. The jev-handoff folder contains the complete skill, runner, five forms and their sample emails. Python 3.10 or later is the only runner dependency.
Copy jev-handoff into ~/.codex/skills/. Open a new task and invoke $jev-handoff.
Copy jev-handoff into ~/.claude/skills/. Open a new session and invoke /jev-handoff.
Set OPENROUTER_API_KEY in the environment used by your agent or terminal. The runner reads it there. Keep the key out of prompts, screenshots and shared files.
This works in an agent with terminal and file access. ChatGPT and Claude web chats need a separate action, connector or MCP connection. The skill does not install Jev inside a model.
Start with the skill command for your agent, then paste this:
Classify the 20 synthetic spam emails included in jev-handoff.
Validate the form with a dry run, then call Jev using my loaded
OpenRouter key. Keep full results on disk. Show total time,
reported cost, errors and uncertain cases.
Take no action on the messages.
That authorizes one sample of 20 API requests. A dry run makes no paid requests. The default is four workers with a 60-second timeout per request.