The Agent Skills Directory

[SAFE]: The skill implements a robust, multi-phase production-prep workflow including narrative analysis, bible extraction, and reference image generation. There are no signs of malicious code, data exfiltration patterns, or obfuscated content across the 11 files analyzed.
[COMMAND_EXECUTION]: The skill explicitly instructs the agent to run all image generation and consistency verification phases 'silently' without user confirmation gates. While this streamlines the asset pipeline, it represents an increase in agent autonomy and a reduction in manual oversight for external tool interactions.
[PROMPT_INJECTION]:
Ingestion points: The skill processes untrusted narrative source material such as scripts, stories, and prose provided by the user (SKILL.md).
Boundary markers: The workflow uses a 'Prompt Keyword Library' and 'Style Anchors' to enforce visual consistency, which partially constrains the impact of the source material on the final output (Phase 2.4, Phase 11).
Capability inventory: The skill uses the 'nanobanana' image-generation MCP for tool calls including generate_image, edit_image, and character_consistency (SKILL.md, references/reference-image-guide.md).
Sanitization: Style prompts are constructed using a predefined vocabulary library, but the narrative content itself is interpolated into action descriptions without explicit sanitization against embedded instructions (references/prompt-keyword-library.md).
Risk: This creates a surface for Indirect Prompt Injection where malicious content within a script could attempt to influence the image generation tool's parameters or output.

scene-inventory-extractor