The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata and data samples from external systems, creating a surface for indirect prompt injection.
Ingestion points: Data enters the context through tools like discover_connection_schema and preview_pipeline (detailed in references/product-capabilities.md).
Boundary markers: The skill contains instructions to restrict preview data to the current request and to use curated error hints rather than full logs.
Capability inventory: The agent has permissions to perform operational tasks such as start_pipeline, update_pipeline_draft, and stop_run.
Sanitization: Guidelines enforce strict adherence to tool schemas and stable idempotency keys, though content-level sanitization of external data is not explicitly detailed.
[DATA_EXFILTRATION]: The skill connects to the joinlayer.app infrastructure. It prioritizes data safety by prohibiting the collection or transmission of credentials within the AI conversation. Authentication is handled via secure OAuth flows in the user's browser, preventing the agent from seeing sensitive tokens or secrets.
Evidence found in SKILL.md and references/connector-contracts.md.
[EXTERNAL_DOWNLOADS]: The skill references and connects to the vendor's official MCP endpoint (https://mcp.joinlayer.app/mcp). This is a legitimate external dependency required for the skill's primary operation and is within the expected scope of the service provider.
Evidence found in agents/openai.yaml and SKILL.md.

joinlayer-pipelines