Skip to content
Swap Meet

Free

No. 032Free

adversarial-research: sort claims the sources support from the rest

Sends separate verifier agents after each claim and labels it survives, cut or unconfirmed. A run costs about 200 agents and 20 minutes.

by Lucas Cooper-Bey

About it

adversarial-research gathers claims, then sends separate agents to try to refute each one. Each claim is labeled survives, cut or unconfirmed, with the link that settled it. An HTML report groups them. High-stakes claims can use three verifiers and a majority. Thin evidence stays unconfirmed. The skill says a run costs about 200 agents and 20 minutes. The idea came from Mark Kashef's Dynamic Workflows prompt pack. Limits: A run costs about 200 agents and 20 minutes, and the skill says to scope the depth. Verifiers must cite the source that settled a claim. Thin evidence stays unconfirmed. Forcing a schema on the verifier agents dropped most results on one recorded run; take a one-line verdict instead. A supported claim is only as good as the sources found. You need research tools, or Claude Code's web search and fetch. Door 33 of 100 Doors.