The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data from /mnt/agents/upload/ (such as .docx, .pdf, .txt, and .md files) to extract plot, characters, and style information. While the workflow emphasizes summarization and indexing to manage context, there is an inherent attack surface where instructions embedded in these source files could potentially influence the agent's behavior during subsequent writing stages.
Ingestion points: User-provided reference files in /mnt/agents/upload/.
Boundary markers: The skill uses a structured multi-stage workflow (Intake -> Outline -> Draft) and emphasizes summarizing source material into separate 'source_notes' files, which provides a degree of separation.
Capability inventory: The skill has capabilities to read/write files in /mnt/agents/output/, execute shell commands (ls, wc, grep, sed), and run Python code via ipython and local scripts.
Sanitization: The provided analysis scripts use regex for counting and pattern detection but do not perform explicit sanitization or filtering of instruction-like content in the source text.
[COMMAND_EXECUTION]: The skill uses standard shell utilities including ls, wc, grep, and sed for file system navigation and basic text analysis. These tools are used for legitimate project management tasks such as file listing and structural mapping of chapters.
[DYNAMIC_EXECUTION]: The skill executes local Python scripts (check_wordcount.py, scan_anti_ai.py) for word counting and stylistic analysis. It also leverages an interactive Python environment (ipython) for structural analysis of documents (e.g., using python-docx for paragraph counts). These executions are limited to local file processing and do not involve remote code retrieval.

creative-writing