The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references external resources from non-trusted domains to define its design rules and visual logic.
Evidence: Multiple links to a personal GitHub repository (github.com/freestylefly/awesome-gpt-image-2) are included in references/cinematic-image-direction.md as sources for visual case adaptations.
Evidence: Links to specific video IDs on Douyin (e.g., douyin.com/video/7661966023750405382) are listed in references/SOURCE-LEDGER.md as sources for lighting and scene scheduling methods.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data which is then interpolated into complex prompt templates, creating a vulnerability to indirect injection.
Ingestion points: The skill explicitly mentions using "user-supplied images" (10 reference images) as the basis for the "cinematic image direction" and "source ledger" (found in references/SOURCE-LEDGER.md and references/cinematic-image-direction.md).
Boundary markers: While the skill uses structured 10-section formatting, it lacks clear boundary markers or instructions to the agent to ignore any text-based instructions potentially embedded in the source images or descriptions.
Capability inventory: The skill generates a comprehensive style_route containing camera, space, action, and material parameters used to drive downstream AI image and video generation tools.
Sanitization: No sanitization, filtering, or validation steps are defined for the data extracted from the reference images or user descriptions before they are used to populate the design templates.

cyberpunk-design