Click-and-run adversarial research that cuts past marketing and SEO. Ask a plain question about a product, tool, supplement, claim, or study and get back a calibrated verdict — marketing, or the real thing — built by a six-member council that argues with itself instead of reproducing the first vendor-friendly result. It does not flatter you, and it does not reflexively deny everything.
curl -fsSL https://bneck.com/research | bashInstalls the skill for both Claude Code (~/.claude) and Codex (~/.codex). Then just ask:
- Claude Code:
/research-armed— or plainly, "is X marketing or real?" - Codex:
$research-armed— or "stress-test these claims" - Web (Claude.ai / ChatGPT): drop
~/.research-armed/dist/research-armed.mdinto a chat, then ask
No flags, no modes, no setup beyond the line above. You supply the subject; the skill does the scoping, searching, debate, and verdict, and picks how deep to go on its own.
A pre-flight Scoper neutralizes your question, then four research vectors work in parallel: a Prosecutor and a Defender with symmetric burdens (popularity is evidence for neither side), a Field Investigator that leaves the SEO'd surface for forums, complaint boards, and teardowns, and a neutral Source Grader that scores every source by independence and flags marketing machinery. They cross-examine — each must concede a point and rebut the strongest opposing one — and a separated Adjudicator returns one of five verdict levels, from pure marketing to genuinely substantiated, with a real "real core wrapped in marketing" middle. The verdict is anchored to an evidence ledger and guarded against both flattery and reflexive denial.
See master.md for per-platform manual install, the build pipeline, and the protocol source
in core/. The protocol is authored once in core/; build.sh fans it out to all three builds.