Quicksilver
A Claude Code skill that hands bulk scanning to Jev so Claude reads a shortlist instead of everything.
A Claude Code skill that hands bulk scanning to Jev so Claude reads a shortlist instead of everything.
A user task like "which of these 187 files handle auth" becomes a fixed candidate set in the state and one typed question per candidate, sent in parallel to Jev. The labels, yes/no answers or scores come back with their confidence, and only the shortlist enters Claude's context, so the expensive model reads four paths instead of 187 files.
Is this file part of the authentication flow the task asks about?
The state carries one candidate file, its path and a slice of its content, and the task being answered. The documented benchmark sends 187 Hono files for one discovery task; three of the question set are shown here.
Is this file part of the authentication flow the task asks about?
The state carries one candidate file, its path and a slice of its content, and the task being answered. The documented benchmark sends 187 Hono files for one discovery task; three of the question set are shown here.
Is this file part of the authentication flow the task asks about?
The state carries one candidate file, its path and a slice of its content, and the task being answered. The documented benchmark sends 187 Hono files for one discovery task; three of the question set are shown here.
Claude uses the skill when a task means reading a lot to decide a little, such as finding files, triaging logs or routing tickets.
The skill collects the pile to judge, from a directory listing, a log file or a ticket batch, and keeps each item's identity in the state.
One typed question per candidate goes to Jev with the task text, and the answers return with their confidence.
Only the items that pass enter Claude's context, with their confidence, so the model spends tokens on the answer rather than the pile.
It turns context waste into a bounded decision problem with a measured trade: the published benchmark cuts Claude's tokens by a median 82% and matches its accuracy on eight of twelve real tasks, while two log-triage cases show the accuracy actually falling when the cheap shortlist misses.