Geoffrey Litt outlines explanation artifacts, quizzes, and micro-worlds to maintain human participation in agent-driven codebases

Litt presented three education-derived techniques for rapid understanding: auto-generated explain-diff documents that supply background context and intuition before code changes, self-administered quizzes that surface gaps in mental models, and executable micro-worlds that allow isolated experimentation. The /explain-diff skill he ships outputs HTML or Notion pages that first describe the prior system state, then the goal, then the delta. These artifacts are already used daily inside Notion teams.

The argument rests on iterative loop dynamics rather than single-shot verification. Once agents handle correctness checks, the remaining human bottleneck is the ability to generate the next productive prompt. Litt links this to cognitive debt concepts from Margaret Storey and Simon Willison: short-term omission of understanding compounds into reduced fluency across subsequent cycles. Data from his internal usage shows explainers reduce time to form follow-up ideas compared with raw diffs.

Operationally this shifts tooling priorities from pure generation speed toward persistent, queryable representations of system state. Teams that treat understanding as a first-class output can sustain higher conceptual bandwidth across multi-week agent collaborations. The pattern mirrors historical shifts from assembly listings to higher-level abstractions when compiler output became reliable.

Next steps include wider deployment of quiz generators and micro-world sandboxes tied directly to agent sessions, with measurable thresholds on iteration rate per developer hour.