Tip: Quicksilver offloads Claude Code's 'read a lot to decide a little' calls to Jev, cutting tokens 86% but losing accuracy on log triage
GitHub·low signal
The skill, created 09-25, sends bulk yes/no, label and score calls to TypeSafe's Jev at $0.042 per million tokens and returns a shortlist to Claude. Its own 12-task benchmark matched Claude alone on 8 tasks, including 187-file codebase discovery (26.4k to 2.4k tokens) and 'where is X' ranking. Log-triage F1 fell from 54% to 23% and a security-review shortlist from 100% to 89%. Use it for search and routing, not for finding failures.