The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content (text, files, or URLs) which is then interpolated into prompts for the image_generate tool. An attacker could potentially embed instructions in the source content to manipulate the visual output or bypass design constraints.\n- Ingestion points: User-provided content is saved to source.md and processed into structured-content.md (Step 1 and 2 in SKILL.md).\n- Boundary markers: The skill uses a structured template for content organization but lacks explicit boundary delimiters or instructions to ignore embedded commands within the final prompt template (references/base-prompt.md).\n- Capability inventory: The skill utilizes write_file, read_file, and image_generate.\n- Sanitization: While the skill includes proactive instructions to strip credentials and secrets from output (Step 2 and Pitfalls), it does not explicitly sanitize for prompt injection patterns.

baoyu-infographic