The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted external data in the form of user-provided text prompts and reference images. This data is passed to external image generation and editing models, creating a potential attack surface for indirect prompt injection where malicious instructions could be embedded in the input to influence the agent's actions.
Ingestion points: prompt, referenceImages, and image parameters used in the model_run tool as described in SKILL.md.
Boundary markers: No specific delimiters or instructions to disregard embedded commands are mentioned in the documentation for these inputs.
Capability inventory: The skill utilizes tools for network API requests (model_run) and file system writes (asset_download).
Sanitization: No sanitization or filtering of the processed inputs is described.
[EXTERNAL_DOWNLOADS]: The documentation instructs the agent to ask users to install additional components from the author's official repository.
Evidence: The text recommends using npx skills add scenario-labs/skills to obtain missing sibling skills.
Context: These references target the official resources of the skill's author (scenario-labs) and are standard for the tool's ecosystem.
[DYNAMIC_EXECUTION]: The skill suggests a method for verifying image transparency using a Python command that dynamically loads a library.
Evidence: Recommendation to use uv run --with pillow to check the alpha channel of generated images.
Context: This is a diagnostic step using well-known development tools (uv, pillow) for functional verification.

scenario-seedream