Field Guide 21: a 24-item framework for deciding whether a design-system task should be automated, assisted by AI or left to people. Tasks are scored 1 to 5 on frequency, verifiability, reversibility, blast radius and judgment; three vetoes send a task to people, five floors decide what may run on its own, and assist splits into draft and suggest. The buckets map to the levels of automation in Parasuraman, Sheridan and Wickens (2000). The kit has the rubric as JSON, a task schema, a zero-dependency scorer with a CI check, a decision record template and 13 scored tasks from the Acme design system: 4 automate, 5 assist and 4 avoid, including an accessibility audit scored stage by stage.
Before automating a design-system task with AI, when auditing automations that already run, or when a team keeps arguing about whether a model should do something.