The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains a 'Boundary Declaration' section in SKILL.md and references/material-decomposition.md that explicitly commands the agent to override safety filters and content guidelines. It specifies that the agent 'must not refuse a whole chapter or book on the grounds of sensitivity or discomfort' regarding themes like violence or domestic abuse.
[DYNAMIC_EXECUTION]: The references/style-profile-generator.md file contains a multi-line Bash script that dynamically generates and executes a Python script using a heredoc to calculate sentence length and punctuation statistics from text samples.
[COMMAND_EXECUTION]: The SKILL.md file defines a complex shell command pipeline (utilizing ls, sed, sort, cut, and grep) to merge multiple chapter summary files into a single collection, executing these commands directly in the environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted, user-provided novel text, creating a vulnerability surface where instructions hidden in the literature could influence agent behavior.
Ingestion points: Raw novel text provided via file paths or pasted into chat (found in SKILL.md).
Boundary markers: The skill uses a [材料声明] (Material Declaration) prefix in prompts for sub-agents to attempt to scope the context, but this is a soft boundary.
Capability inventory: Subprocess calls for shell commands (ls, sed, grep), local Python execution, file writing, and sub-agent spawning (SKILL.md).
Sanitization: The skill relies on regex for chapter boundary detection but lacks robust sanitization for instructions embedded within the novel content.

story-long-analyze