Best Practices: One-Shot

A one-shot asks N seats the same question concurrently and returns N independent answers. The human reads the answers; nothing synthesizes them. It is the correct mode when the deliverable is perspectives, not a decision.

Core Rule

Never manufacture agreement from one-shot answers.

The moment you summarize the answers into "the models agree that..." you have run a roundtable without its equal-packet, quorum, and attributed-dissent guarantees. The moment you pick "the best answer" you have run a competition without isolation or an independent judge. If the user wants either, recompile in the right mode; do not upgrade a one-shot in prose.

Contract

skills/ask/run.sh one-shot "<question>" --handler <a> --handler <b> ...