The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external JSONL payloads and fetches live documentation from doubleword.ai and docs.doubleword.ai, creating a surface for potential injection of malicious instructions via untrusted data.
Ingestion points: Local JSONL data files, external product/model indexes (llms.txt), and markdown documentation pages fetched at runtime.
Boundary markers: The skill instructions explicitly require the agent to validate JSONL files locally using dw files validate and dw files stats before any upload or processing occurs.
Capability inventory: Terminal execution for dw CLI commands, Python code execution via the autobatcher library, and network requests via curl to the vendor's API.
Sanitization: Relies on the dw CLI's built-in validation tools to verify payload structure and integrity before transmission.
[COMMAND_EXECUTION]: The skill executes shell commands via the dw CLI and curl to perform inference and manage data jobs.
Evidence: Includes recipes for dw realtime, dw batches create, and dw stream. These commands are directed towards the vendor's official services.
[EXTERNAL_DOWNLOADS]: The skill fetches documentation and model metadata from the vendor's infrastructure.
Evidence: Instructions include fetching from https://doubleword.ai/llms.txt and https://docs.doubleword.ai/ to maintain up-to-date model information.

doubleword