The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (such as project documentation or source code) and feeds it to an LLM via the claude CLI, which creates an attack surface for indirect instructions.\n- Ingestion points: File content passed via the --files flag or direct prompt interpolation in the claude command.\n- Boundary markers: Absent. The skill documentation does not define explicit instruction delimiters or "ignore embedded instructions" warnings to separate user instructions from processed file content.\n- Capability inventory: Executes shell commands via claude, a local script (delegation_executor.py), and make automation.\n- Sanitization: Absent. No sanitization or validation of input file content is mentioned before it is processed by the model.\n- [EXTERNAL_DOWNLOADS]: The skill instructs users to install the official @anthropic-ai/claude-code package from the NPM registry. This targets a well-known package from a trusted organization (Anthropic) and is documented neutrally as a prerequisite.\n- [COMMAND_EXECUTION]: The skill utilizes several command-line tools, including claude, python, and make, to facilitate its core delegation functionality. These commands are configured via environment variables and do not involve arbitrary shell execution or unsafe user-supplied arguments outside of their intended operational context.

glm-delegation