The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains "Boundary Declarations" and "Material Declarations" in SKILL.md that explicitly instruct the agent and its sub-agents to ignore safety filters. It specifically commands the AI to extract details of violence, domestic abuse, and "dark ethics" as "regular fictional narrative elements" and strictly forbids refusing chapters or books based on sensitivity or discomfort. This is a deliberate attempt to override the underlying AI safety protocols.
[COMMAND_EXECUTION]: The skill performs local command execution. It runs a Python script (scripts/render_chapter_summary.py) to validate and render data. Additionally, references/style-profile-generator.md contains complex shell pipelines that execute dynamic Python one-liners to calculate text statistics (sentence length, punctuation density) on user-provided content. While these are part of the skill package, they represent a significant execution capability within the local environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted novel text provided by users. It lacks robust boundary delimiters when interpolating this text into prompts for sub-agents (using only simple labels like 章节原文：\n{原文文本}). Combined with the explicit instructions to bypass safety filters, this creates a vulnerability where malicious instructions embedded in the novel text could influence the agent's behavior or output generation.
Ingestion points: Reads novel text files and user-pasted text as defined in SKILL.md Phase 1.
Boundary markers: Missing robust delimiters; relies on simple newline separation.
Capability inventory: File system writing, sub-agent spawning, local script execution, and shell command execution.
Sanitization: Employs Markdown character escaping in the render_chapter_summary.py script.

story-long-analyze