The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to orchestrate communication between multiple AI agents, creating an inherent surface for indirect prompt injection where instructions could be embedded in the responses of coordinated roles.
Ingestion points: Data is ingested from the [RESPONSE:role] blocks of other agents, the ~/.claude/decisions/ persistent storage, and local context files like company-context.md or decision-log.
Boundary markers: The protocol employs structural delimiters such as [INVOKE:role|question] and [RESPONSE:role] to separate agent inputs, and requires explicit tagging for assumptions using [ASSUMPTION: ...].
Capability inventory: The skill possesses the capability to read and write to the local file system within the ~/.claude/ directory and to initiate cross-agent invocations.
Sanitization: The skill relies on logical "Self-Verification" and "Peer Verification" checklists to ensure data accuracy and attribution, but lacks programmatic sanitization or filtering of the raw text strings received from other agents.
[NO_CODE]: The skill consists entirely of markdown configuration and instructional guidelines. It does not contain any Python scripts, Node.js modules, or shell executables, which significantly reduces the direct execution risk profile.

agent-protocol