Context pointer

A context pointer is a one-line reference to information rather than the information: “the auth flow is described in docs/auth.md,” or “run pnpm test:e2e to see the browser tests.” It occupies a few tokens in the context window and tells the agent that a resource exists, where it is, and when it is relevant. When a task calls for it, the agent chases the pointer via a tool call; otherwise it leaves it alone.

Pointers are the mechanism behind progressive disclosure, and their absence is why instruction files bloat. Without a pointer, the only way to make sure the agent knows something is to paste it, so people paste. With a pointer, the knowledge costs almost nothing until it is used. The failure mode of pointers is vagueness: “see the docs” points nowhere, and the agent will not go looking.

Write pointers with a path, a trigger, and a reason. Path: the exact file, command, or URL. Trigger: the situation that should send the agent there. Reason: one clause on why it matters, so the agent can judge relevance. Keep them in AGENTS.md for project-wide facts and in handoff notes for task-specific ones. The primary source stays where it is; only the pointer travels.