sami-mag07/claude-skill-research-agent
41 stars · Last commit 2026-10-02
Research skill and researcher subagent for Claude Code: multi-engine search behind one interface, parallel agent waves, and a verify gate that blocks any citation without a verbatim quote. No personal data; configured via .env.
README preview
# claude-skill-research-agent [](https://github.com/sami-mag07/claude-skill-research-agent/actions/workflows/tests.yml)    A research skill for Claude Code that treats web research like engineering: several search engines behind one interface, parallel agent waves that never duplicate each other's work, and a verify gate that refuses to call a report done while any citation lacks a verbatim quote from its source. Zero runtime dependencies. Every script is Python stdlib, so there is nothing to `pip install` and nothing to audit beyond the files in this repo. ## Why Agent-driven research fails in three predictable ways: 1. **One search backend.** The harness search tool hits a shared quota, a rate limit, or simply misses the niche page. The agent then concludes "nothing exists". 2. **Parallel agents step on each other.** Five subagents on one topic return the same ten URLs five times. 3. **Citations that do not hold.** The report cites `[3]`, but the sentence it supports is nowhere on page 3. Each script here removes one of those failure modes.