The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The pipeline architecture described in SKILL.md and implemented in scripts/pipeline_template.py processes untrusted external data, introducing a surface for indirect prompt injection. \n- Ingestion points: Data is acquired from external sources such as APIs and web content in the acquire stage (found in SKILL.md and scripts/pipeline_template.py). \n- Boundary markers: The skill suggests using section headers but does not provide specific implementation for robust delimiters to separate untrusted content from system instructions. \n- Capability inventory: The provided scripts demonstrate file system access (writing and deleting state files) and network calls via LLM APIs. \n- Sanitization: The instructions do not detail methods for sanitizing or escaping ingested data before it is used in prompt templates during the prepare stage.

project-development