When you ask Claude or Codex to fix a typo on your website, it may also read the standing instructions you saved for quite different jobs. Your conversation, those instructions, opened files and tool results all contribute to the agent’s context, the material it has available while it works. The context window limits how much it can consider at once.

Unrelated instructions add reading without helping the task. Long context can also make relevant details harder for a model to use reliably. You still want your email guidance available when you ask for an email; the typo fix simply has no use for it.

Routing gives the agent a short directory of its detailed instructions. It reads that directory first and follows the relevant links as the work calls for them. This course shows how to organize those routes and check that the needed guidance is actually read.