The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data retrieved from a knowledge base (referred to as the "brain") which may contain information from various sources (e.g., wiki, gstack, yc-media). This content could include malicious instructions designed to manipulate the agent's behavior.
Ingestion points: Information enters the agent context through tools such as recall, search, query, and get_page, as well as CLI commands like gbrain query and gbrain think.
Boundary markers: The instructions mention using ## Facts fences and [Source: ...] citations for organization, but these do not serve as cryptographic or robust security boundaries against instruction parsing.
Capability inventory: The skill possesses significant write capabilities, including put_page, add_link, and add_timeline_entry, and the ability to execute shell commands via the gbrain CLI.
Sanitization: The skill implements automated secret scanning and redacts high-entropy values like credentials (e.g., <REDACTED:url_credentials>), though it does not explicitly detail sanitization for prompt injection within the text body.
[COMMAND_EXECUTION]: The skill relies on the execution of the gbrain CLI tool to perform its core functions.
Evidence: The SKILL.md file contains numerous examples of shell commands, such as gbrain recall --query ... --json, gbrain sources list, and gbrain think "...".
Context: These commands are presented as the standard operational protocol for interacting with the knowledge base infrastructure provided by the vendor.

brain-ops