The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (e.g., articles, user queries) by interpolating them into prompts for structured extraction.\n
Ingestion points: Functions such as extract_person, structured_extraction, and agent_workflow in SKILL.md accept strings from external sources.\n
Boundary markers: The skill leverages SGLang's constrained decoding capabilities (json_schema, regex, and grammar in sgl.gen()) which enforce strict output formatting, acting as a boundary to prevent the model from following embedded malicious instructions.\n
Capability inventory: Analysis of the skill's scripts (SKILL.md) shows the tool is used for API-based inference; no patterns for arbitrary file system access or system command execution using this data were found.\n
Sanitization: The skill does not perform explicit sanitization or escaping of the input variables before prompt interpolation.\n- [EXTERNAL_DOWNLOADS]: The skill instructs the user to install the sglang framework and hardware-optimized dependencies from public registries.\n
Evidence: SKILL.md includes commands for pip install "sglang[all]" and git clone https://github.com/sgl-project/sglang.git. It also points to the official FlashInfer wheel index at https://flashinfer.ai/whl/cu121/torch2.4/.\n- [COMMAND_EXECUTION]: The documentation provides shell commands to launch and manage inference servers and deployment infrastructure.\n
Evidence: SKILL.md and references/deployment.md contain commands for python -m sglang.launch_server, docker run, and Kubernetes manifest examples.

sglang