The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for 'Query Expansion' and 'Query Understanding' that interpolate raw user input directly into LLM prompts, creating a surface for injection attacks.
Ingestion points: The query parameter in the expand_query function (found in references/patterns.md) and the understand_query function (found in references/sharp_edges.md).
Boundary markers: No specific delimiters (like XML tags or triple quotes) or 'ignore' instructions are used to separate the user-provided query from the prompt instructions.
Capability inventory: The functions utilize llm.complete to generate query variations or extract entities, which could be manipulated to execute unintended instructions if the input is adversarial.
Sanitization: The code snippets do not include logic for sanitizing, escaping, or validating the query string before it is interpolated into the f-string prompt templates.

vector-specialist