The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (images and text descriptions) which is then used to generate a prompt for an image generation tool.
Ingestion points: In SKILL.md (Workflow Step 1), the agent is instructed to analyze user-provided images and text to identify subjects and traits.
Boundary markers: The skill utilizes a structured references/avatar-prompt-blueprint.md as a template. It uses labeled fields (e.g., Subject identity, Recognizable traits) which serve as delimiters for user-provided data.
Capability inventory: The skill calls an image generation tool (image_gen) and performs file system write operations to save prompts and images in the punk-assets/ directory.
Sanitization: The instructions require the agent to extract "concise visual traits" and "derived subject fields" rather than performing raw interpolation of user strings, which provides a functional layer of sanitization and reduces the likelihood of the agent obeying malicious instructions embedded in the input.

punk-avatar