Store reusable code as scripts inside skills (plus append-only state) so the agent composes instead of regenerating boilerplate

Anthropic/Claude·medium signal

Anthropic recommends storing reusable code as scripts inside a skill rather than as instructions describing how to write that code, so the agent composes and decides rather than regenerating boilerplate each run — cutting token spend and output variance. Pair this with append-only logs or JSON in ${CLAUDE_PLUGIN_DATA} so the skill carries state across sessions and can do delta updates instead of recomputing from scratch. It turns the file system itself into a progressive-disclosure context layer.