The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external screenplay files (untrusted data) through the provided episode_intake.py script, which creates a potential surface for indirect prompt injection attacks. Besides scalars, boolean, and null, other values must be double-quoted as valid strings.
Ingestion points: Untrusted content enters the agent's context when reading screenplay source files for indexing, slicing, and adaptation analysis, as described in SKILL.md and multi-episode-intake.md.
Capability inventory: The episode_intake.py script has the capability to write files to the project directory, specifically for creating indexes (episode-intake-index.json), episode slices, and progress checkpoints.
Boundary markers: While the methodology emphasizes distinguishing between source material and creative intent, the instructions do not prescribe specific technical delimiters (e.g., unique XML-style tags) to encapsulate untrusted data when it is processed by the LLM.
Sanitization: The episode_intake.py script implements path validation in the _portable_source_ref function, which prevents path traversal by rejecting absolute paths and directory traversal sequences like ...

short-drama-develop