The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains standard shell commands for image optimization using utilities like cwebp, mogrify, and jpegoptim. These are documented for legitimate image processing tasks and include instructions to verify tool availability before execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest local marketing context files and inspect external URLs for image references, creating a surface for indirect prompt injection. Ingestion points: Reads context from .agents/product-marketing.md, .claude/product-marketing.md, and product-marketing-context.md; fetches image metadata from external URLs. Boundary markers: Explicitly commands the agent to refuse connections to loopback, link-local, or private-network destinations. Capability inventory: Includes file reading, shell command execution, and read-only network access via browser or HTTP tools. Sanitization: Strictly prohibits attaching cookies or credentials and requires verification of public HTTPS URLs.
[DATA_EXFILTRATION]: The skill includes explicit security boundaries to prevent data exfiltration when inspecting external URLs, specifically forbidding the transmission of local files or workspace content.

suede-image