The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data in the form of uploaded photos and natural language descriptions to generate Riso-style portraits. This content is interpolated into prompts for the gpt-image-2 model, creating a surface for indirect instructions.
Ingestion points: User-supplied single-person photos and optional art direction preferences (referenced in SKILL.md and references/riso-art-direction.md).
Boundary markers: The workflow uses a template where user notes are appended as a Creator note and are explicitly forbidden from overriding core identity, anatomy, or safety constraints (references/riso-art-direction.md).
Capability inventory: The skill utilizes gpt-image-2 for raster image editing and executes a local script (scripts/profile_store.py) to write to the file system.
Sanitization: The profile persistence script includes a blacklist (SENSITIVE_PARTS) to prevent the storage of sensitive fields such as tokens, secrets, and API keys.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute local Python scripts for state management and source validation. Specifically, scripts/profile_store.py is used to read and write user-owned profile data (e.g., color palettes and cropping preferences) to a local JSON file located in the user's home directory or config directory. While these scripts are designed for legitimate preference management, they represent an execution capability on the local host. (References: SKILL.md, scripts/profile_store.py, scripts/validate_skill.py).

lov-riso-portrait