The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external media (URLs and file uploads) and structured API responses, which constitutes an indirect prompt injection surface.
Ingestion points: Media content enters the agent's workflow via public HTTPS URLs or local file uploads to the POST /detect and POST /intelligence endpoints. The resulting JSON responses are then parsed by the agent to determine media authenticity.
Boundary markers: The skill provides strong boundary markers, instructing the agent to never speculate on authenticity and only rely on completed detection results using specific response fields like item.metrics.label and aggregated_score.
Capability inventory: The skill uses curl for network communication with the vendor's API and utilizes python3 or jq for parsing structured JSON data in a controlled environment.
Sanitization: Instructions require the agent to interpret scores within defined ranges (e.g., 0.0–0.3 for authentic media) and include operational details in the output, which helps mitigate the risk of the agent misinterpreting potentially poisoned media metadata or API responses.

resemble-detect