The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted upstream data (scripts, characters, and art assets) and interpolates them into prompts intended for other AI models (MiniMax H3 and Seedance).
Ingestion points: The script reads script.json, outline.json, cast.json, and art.json from the local file system via CLI arguments.
Boundary markers: The skill wraps dialogue and metadata in protocol-specific delimiters such as <d>[Chinese] ...</d> for H3 and {} for Seedance to maintain structure.
Capability inventory: The skill utilizes the node:fs module for reading input files and writing output (storyboard JSON, Markdown, HTML reports, and log files). It does not have network access or subprocess execution capabilities.
Sanitization: The tool implements 18 deterministic 'quality gates' in scripts/novel-storyboard.mjs that validate data coverage, duration constraints, and naming conventions. It also employs an esc() function to sanitize data before embedding it into HTML reports, preventing potential cross-site scripting (XSS) in the generated artifacts.
[SAFE]: The skill is self-contained and adheres to a 'zero-dependency' policy, using only Node.js standard libraries. Analysis of the scripts (novel-storyboard.mjs and selftest.mjs) confirms they do not contain malicious code, remote downloads, or unauthorized data exfiltration patterns. The use of local storage for logging validation results (.gates.jsonl) is a standard utility function and is correctly excluded from version control via .gitignore.

novel-storyboard