Superhunter — Development skill for Claude Code

SuperHunter

A job-search agent built on the Claude Agent SDK. It runs unattended, finds open roles that fit a written profile, verifies each one against the live listing, scores it A to D, and reports over Telegram. It has run every weekday since June 2026.

The interesting part is not the search. It is how the judgment is organised: a deterministic TypeScript orchestrator does everything that can be done without a model, model calls are reserved for the two steps that need judgment, and the rules that drive that judgment live in a versioned file outside the code, with a frozen fallback and a staleness check.

Author: Robin Vorrink.

What it does

Every run is the same five steps:

1. Prep. Load the targeting rules (see the contract below), the source universe (config/sources.yaml), and the set of every URL ever seen, for deduplication.
2. Discover. Fan out 25 discovery agents in parallel. Each agent gets one assignment: a web-search query family, or a batch of six company careers pages. Each returns a JSON list of candidate roles with canonical URLs. Agents may not invent roles or URLs.
3. Dedupe. Drop anything already in the ledgers, by URL and by ATS job id, so the same role behind several links counts once.
4. Verify. For each fresh candidate, fetch the listing page once in Node and make one model call on a stronger model with a hard turn cap. The call confirms the page is live, applies the hard gates, scores A to D, names a rejection bucket, and writes a two-sentence summary for the reader. If the page is walled, escalate once to a browsing agent that reads the company's own ATS page.
5. Track. Append every scored lead to a TSV ledger, surface A and B leads to a markdown tracker, write a per-run digest, append a metrics block, and push the digest to Telegram through n8n.

Drafting application materials is a separate command that runs only when asked (`npm run draft`). Nothing is ever sent on the candidat