The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill allows an agent to resume experiments or check their status, which involves reading the contents of existing markdown files such as plans and checkpoints. These contents are then returned to the agent's context, creating a surface where malicious or accidental instructions in those files could influence the agent's actions.
Ingestion points: The scripts/resume.sh script reads the latest plan and checkpoint files from the experiment directory, and scripts/status.sh reads the first 15 lines of the latest checkpoint.
Boundary markers: The skill wraps the ingested content in <file-content source="..."> XML-like tags to help the agent distinguish it from other instructions.
Capability inventory: The skill possesses capabilities for file system read/write, directory manipulation, and execution of local Bash scripts. It also accesses Git configuration to identify the experiment author.
Sanitization: No sanitization or filtering is applied to the content of the documentation files before it is displayed to the agent.

doc