The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process large volumes of untrusted third-party data such as audio libraries, email takeouts, and API dumps. It includes robust mitigations, instructing the agent to treat all fetched content as data only, use specific boundary markers (untrusted-quoted), and flag agent-directed imperatives during transformation to prevent content from redirecting the pipeline.\n
Ingestion points: External data sources specified in Phase 2 and 3 of SKILL.md.\n
Boundary markers: Explicitly mandates the use of untrusted-quoted fences and untrusted_directives: true flags for all ingested content.\n
Capability inventory: The skill is authorized to write pages to projects/ and sources/ directories and can submit background jobs via the gbrain CLI.\n
Sanitization: Instructions require the agent to identify and isolate potential imperatives within raw data during the transformation phase.\n- [COMMAND_EXECUTION]: The skill documents the use of shell commands via gbrain jobs submit shell for executing background ingestion tasks and periodic git checkpoints. It explicitly identifies this as a high-risk capability and correctly notes that it requires manual environment-level authorization (GBRAIN_ALLOW_SHELL_JOBS=1) by the system operator.\n- [DYNAMIC_EXECUTION]: The skill utilizes dynamic agent invocation (gbrain agent run) and job submission to handle partitioned workloads. These patterns are used for legitimate task orchestration and include guidance on partitioning work to avoid race conditions and state corruption.

bulk-ingestion