The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external document files (e.g., .docx, .pdf, .txt, .md) uploaded by the user, which creates a surface for indirect prompt injection where malicious instructions embedded in the source script could attempt to influence the agent's analysis or output.
Ingestion points: The scripts/split_short_drama_source.py script reads the raw text from provided files for ingestion into the project structure.
Boundary markers: The skill defines a "Three-Layer Control" model (described in references/three-layer-control.md) to separate hard project gates (Foundation) from creative implementation (Skeleton/Flesh), and uses [待确认] placeholders for uncertain data.
Capability inventory: The agent has the capability to read and edit project files, execute subprocesses for PDF conversion, and make MCP tool calls to external services.
Sanitization: The skill employs a deterministic checker (scripts/remake_gate_checker.py) and a rigorous report-based workflow (SIR, RMR, FGR) to validate data integrity before proceeding to script generation.
[COMMAND_EXECUTION]: The scripts/split_short_drama_source.py utility uses subprocess.run to call the pdftotext system command for converting PDF documents into readable text.
Evidence: subprocess.run([pdftotext, str(path), str(out_path)], check=True) in scripts/split_short_drama_source.py.
[DYNAMIC_EXECUTION]: The scripts/split_short_drama_source.py script uses the __import__ function to dynamically load optional PDF processing libraries at runtime.
Evidence: module = __import__(module_name) where module_name is selected from a static list ("pypdf", "PyPDF2").

short-drama-remake