The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions and recipes for querying vector databases (pgvector, Qdrant, Weaviate, Pinecone) using user-provided filters and semantic search. This creates a surface for indirect prompt injection if the data retrieved from the vector store contains malicious instructions intended to influence the agent's behavior.
Ingestion points: Similarity search results and metadata filters as described in SKILL.md and references/engines.md.
Boundary markers: The skill does not provide specific instructions for using delimiters or boundary markers to isolate data retrieved from the database from the agent's instructions.
Capability inventory: The agent interacts with external database engines using SQL and Python client libraries and can execute local scripts.
Sanitization: There is no explicit mention of sanitizing or validating the content retrieved from vector databases before processing it.
[COMMAND_EXECUTION]: The skill includes a local utility script, scripts/verify.sh, and instructs the agent to run it to lint and verify vector store artifacts.
The script uses standard shell utilities such as find and grep to perform heuristic analysis on local files.
Evidence: SKILL.md contains the instruction: "Validate a produced index DDL / collection schema with scripts/verify.sh <artifact-file>."

vector-db