The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external construction documents which constitutes a surface for indirect prompt injection.
Ingestion points: Document content is ingested via the Document class and processed by the add_document method within SKILL.md.
Boundary markers: The query method in SKILL.md builds a context string from search results using simple index labeling (e.g., [Source 1: Title]) but lacks explicit instruction-ignore delimiters to prevent the agent from obeying instructions embedded within the document text.
Capability inventory: The skill includes logic for text chunking, embedding generation (simulated), vector similarity search, and answer generation across SKILL.md and instructions.md.
Sanitization: No sanitization, content filtering, or validation of the ingested document content is performed before the data is interpolated into the agent's context.

rag-construction