Graph Engineering is an open-source library of copy-paste prompts for running AI agents as a graph instead of a queue: split the work into isolated angles, fan out in parallel, merge deterministically, then send isolated skeptics to refute the result before anyone reads it.
The templates
- 00 Your First Graph — one claim, three isolated skeptics, one vote; the kill layer alone, in five minutes
- 01 False-Edge Audit — lays out an existing workflow and finds which "and then"s are fake
- 02 Diamond Research — angles → parallel search → adversarial verification → report with confidence labels
- 03 Adversarial Review — feeds finished conclusions to isolated attackers
- 04 Consultant Roundtable — two-round Delphi with dissent kept in the consensus map
- 05 Issue Tree — MECE decomposition into a tree that dispatches itself
Every template ships with a verbatim transcript of a real run, with what went wrong left in — including a 31-agent run pointed at the repo's own premise. Bilingual (English + Traditional Chinese), MIT licensed, and packaged as a Claude Code plugin so each template becomes a slash command.
It is the companion repo to the essay 〈Agent 左右互搏之術〉 and the working notebook for the next one.