The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data for model training, which constitutes an attack surface for indirect prompt injection. 1. Ingestion points: The skill ingests images, audio, and captions via tools such as upload_asset and train (action: upload_images) as described in SKILL.md and the curation reference. 2. Boundary markers: No explicit boundary markers or delimiters are mentioned to separate training captions from the training logic. 3. Capability inventory: The skill has the capability to create models (model_create), configure training (train), and execute trained models (model_run). 4. Sanitization: No specific sanitization or escaping of external caption content is described in the provided instructions.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing sibling skills using the npx command targeting the scenario-labs/skills repository, which is a standard installation pattern for this ecosystem.
[COMMAND_EXECUTION]: The skill utilizes npx for skill installation and employs tool execution wrappers (scenario_tool_execute_read and scenario_tool_execute_write) to perform operations within the agent environment.

scenario-model-training