The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute local Python scripts included in the skill package (scripts/novel_index.py and scripts/selftest.py) to perform chapter indexing, verification, and sampling. It also references an external tool (project_tool.py) from a related 'core' skill directory. These executions are contained to the local environment and are primary to the skill's functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes arbitrary external text provided by users (novels/scripts) through the 输入/ directory. This creates a surface for indirect prompt injection where malicious instructions embedded in a novel could attempt to influence the agent's analysis behavior.
Ingestion points: Raw text files provided by the user in the 输入/ (Input) directory, accessed during the S0 (Indexing) and S2 (Extraction) stages.
Boundary markers: The skill instructions (e.g., in SKILL.md and references/chapter-extraction.md) explicitly require the agent to use source locators and 'spans' for every claim. It mandates de-quoted function summaries instead of copying source paragraphs (Rule NVA-04), which serves as a grounding mechanism against injection.
Capability inventory: The skill has the capability to execute shell commands (python3) to run the indexing and verification scripts.
Sanitization: There is no explicit sanitization of the novel's prose content; the system relies on the LLM's internal guardrails and the structural requirements of the analysis (e.g., fixed output formats for 'Chapter Extraction') to maintain task integrity.

short-drama-novel-analyze