The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied images to perform portrait editing. While this introduces a potential surface for adversarial content (such as hidden text or instructions within the image data), the skill includes robust mitigation strategies. These include explicit identity-preservation constraints in the prompt templates and a mandatory quality gate to verify the final output.
Ingestion points: User-provided source photos are identified as the edit target in SKILL.md (Step 1).
Boundary markers: Prompt templates in references/prompts.md explicitly label the input as "Image 1 is the edit target and the sole identity reference" to prevent the model from interpreting image content as new instructions.
Capability inventory: The skill utilizes image viewing and generative raster-image editing tools, with no evidence of unnecessary network or filesystem access.
Sanitization: A mandatory quality gate process is defined in references/quality-gate.md to check for identity drift, anatomical artifacts, and unrequested changes before delivery.
[EXTERNAL_DOWNLOADS]: The skill's documentation includes references to external repositories and assets for installation and case studies.
Evidence: README.md provides git clone instructions targeting example.com, and cases/cases.json references image assets hosted on raw.githubusercontent.com. These are used for documentation and distribution purposes and do not involve runtime execution of untrusted code.

lov-professional-portrait