Explicit trigger lists stop LLM agents from over- or under-documenting wikis

A developer at Vellum Labs found that LLM agents instructed to autonomously record important decisions either ask for confirmation too often or log everything indiscriminately. After testing across four wikis over three months using Claude Code, the team identified four specific capture triggers: a choice made between alternatives, a reversal of a prior decision, a human-stated rule not found in code, and a measured result from an experiment. An equally important counterpart list tells the agent what to skip, such as implementation details without trade-offs, temporary bug fixes, and confirmations that changed nothing. The system is enforced through a slash command that first checks whether a trigger applies, and a stop hook that blocks session closure if wiki changes remain unpushed. Vellum Labs has released the capture rules and configuration as a free open-source template and Claude Code plugin on GitHub.