The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to override safety guidelines. Both SKILL.md and references/analyze-workflow.md instruct the agent that it must not refuse to process content based on it being "sensitive" or "uncomfortable" (e.g., violence, domestic abuse, or dark themes) because the material is fictional. Furthermore, the prompt generated for the moshu-chapter-extractor sub-agent explicitly commands it to ignore safety refusals.
[COMMAND_EXECUTION]: The skill relies on executing local scripts bundled with the skill to perform core functions. references/analyze-workflow.md contains logic to invoke scripts/chapter_boundary.py via various Python binary aliases (python3, python, py) and scripts/merge-chapter-summaries.js via node. These scripts interact with the local file system to read novels and write analysis results.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process large volumes of untrusted data in the form of novel text provided by users, which creates an attack surface for indirect prompt injection.
Ingestion points: Untrusted novel content is provided via local file paths or pasted directly into the chat session as raw text.
Boundary markers: The skill attempts to use a [材料声明] (Material Declaration) prefix to provide context to sub-agents, though this marker also contains instructions to ignore safety filters.
Capability inventory: The skill possesses significant capabilities, including creating directories (拆文库/), writing multiple markdown files, and executing Python and JavaScript scripts on the host system.
Sanitization: The skill uses regex-based scripts (scripts/chapter_boundary.py, scripts/check_chapter_summary.py) to parse untrusted text, which could be vulnerable to specially crafted inputs designed to cause errors or influence script logic.

moshu-analyze