The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary text prompts and file contents which are then analyzed by an LLM via the claude binary. This creates a surface where malicious instructions embedded in external files could potentially influence agent behavior.
Ingestion points: The skill ingest data via the PROMPT environment variable and the --files command-line argument as shown in SKILL.md.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are documented for the input data.
Capability inventory: The skill utilizes the claude binary and delegation_executor.py, which have capabilities for reasoning, tool use, and file system interaction.
Sanitization: There is no evidence of sanitization or filtering applied to the external content before it is processed by the model.
[EXTERNAL_DOWNLOADS]: The skill instructions include the installation of the @anthropic-ai/claude-code package from the official NPM registry, which is a well-known service.
[COMMAND_EXECUTION]: The skill executes shell commands to perform delegation, including the claude CLI and a local Python script delegation_executor.py. These commands utilize environment variables to swap API endpoints for redirection.

glm-delegation