cross-critique

TL;DR

Synthesizing several subagents' reports alone makes the main agent the bottleneck: it only sees the tradeoffs it happens to notice. This skill has the authors poke holes in each other's reasoning instead. It assumes round one is done, with proposals made independently, by council, by parallel research subagents or ad hoc. The agent condenses each proposal to its recommendation and reasoning, labels them A, B, C and, where practical, hides who wrote which to reduce bias toward the most confident model. It then asks the same subagents, who keep their context, to weigh each rival with both pros and cons, say whether they'd change their own answer, and rank the options. A thin critique gets a follow-up to its author rather than being dropped. The synthesis weighs evidence rather than votes, treats agreement reached after critique as a strong signal, and surfaces the sharpest objection to each option. The round stays read-only unless the task involves changes.

Summarizes the skill at 8792394.