Repeated corrections can become reusable instructions for an AI collaborator. In Eugene Yan's working method, project facts and completed work live separately from preferences and workflow rules. An annotated index tells the model what each document contains and when it is relevant.
Yan then uses session transcripts to identify recurring corrections and update the relevant instructions or checks. The intended benefit is cumulative: completed work supplies context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →, while feedback improves how later work is approached. This is a documented workflow for carrying lessons between sessions, rather than a measured guarantee that the model will stop repeating mistakes.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
