Jev: Rule Selection for Claude Code

0:39

Sep 19

Jev intelligently selects applicable rules for Claude Code based on prompts and file paths, optimizing context token usage and ensuring relevant rule delivery.

SignalWhat it does: · Picks rules per prompt, and again when Claude is about to change a file.

I put Jev inside Claude Code. Jev reads every prompt and picks which of my rules apply. 12 rules in the project, Claude gets the 1 it needs, and a focused session uses about a quarter of the context tokens. Watch: my prompt never mentions payments, yet the payments rule lands the moment Claude edits a checkout file. 40 seconds, real session. The problem: rule files pile up. Test the payment code. Follow the deploy checklist. Use our commit style. Load them all and Claude wades through noise on every prompt. Match them by keyword and a rule is missed whenever the trigger word is absent. jev-rules swaps matching for judgment. One yes/no question per rule, "is this request about that?", all in a single Jev call. In my recordings Jev judged 12 rules in 367 to 773 ms, about $0.00002 per prompt. This is exactly what Jev was built for. What it does: · Picks rules per prompt, and again when Claude is about to change a file. Jev only sees the file path, never the code. · Delivers each rule once per session. Measured over 8 checkout prompts: about 330 tokens, against about 1,390 for loading all 12 rules up front. Touch every topic in one session and it breaks even. It never costs more. · Filters a codebase map the same way, so Claude gets the one doc that answers the question. · Fails open. No key, timeout, Jev down: every rule is injected and nothing is ever blocked. Install and repo 👇 @typesafeai @ClaudeDevs

Original post