The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements search and retrieval functionality which consumes external data (embeddings and metadata filters) that could potentially contain malicious instructions if the upstream data source is compromised.
Ingestion points: query_embedding and filter values (e.g., tenant_id, doc_type) processed in SKILL.md via the qdrant-client.
Boundary markers: The provided snippets do not define explicit boundary markers or instructions to ignore embedded commands in retrieval results.
Capability inventory: The skill manages database operations, including network operations via qdrant-client, shell command execution for Docker management, and SQL execution via PostgreSQL.
Sanitization: No explicit sanitization or validation of the retrieved content is shown, which is typical for operational documentation but represents a surface for indirect injection.
[CREDENTIALS_UNSAFE]: The pgvector Docker deployment example uses a hardcoded placeholder for the database password.
Evidence: docker run ... -e POSTGRES_PASSWORD=secret ... in SKILL.md.
Note: This is documented as a common placeholder for local development and documentation purposes rather than an active credential exfiltration attempt.

vector-database-ops