The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-curated datasets, specifically asset descriptions and auto-generated captions, during the review and curation phases (e.g., via model_get and train tools). These text fields act as ingestion points for external data that the agent reads and reviews.
Ingestion points: Asset descriptions and automaticCaptioning metadata fields read from the Scenario catalog.
Boundary markers: No explicit delimiters are specified for separating asset metadata from agent instructions in the provided documentation.
Capability inventory: Capabilities include file/asset updates (asset_update), training configuration (train), and model inference (model_run).
Sanitization: No specific sanitization or filtering logic is mentioned for processed captions.
Risk assessment: This represents a standard operational surface for a dataset-management skill. Since the data is typically authored by the user or an automated captioner for the purpose of training, the risk of adversarial injection is negligible.
[EXTERNAL_DOWNLOADS]: The instructions advise the agent to suggest the installation of missing sibling skills using npx skills add scenario-labs/skills. This command targets the authorized vendor's own repository, representing expected modular behavior for the platform Ecosystem.

scenario-model-training