The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains 'CRITICAL COMPLETION POLICY' and 'CONTEXT WINDOW' instructions that explicitly direct the agent to 'NEVER ask "Would you like me to continue?"' and to 'continue working indefinitely.' These directives are designed to override standard agent behavior constraints and interaction thresholds to ensure the completion of long documents.
[COMMAND_EXECUTION]: The workflow requires the agent to execute shell commands for document processing and verification. Specifically, it uses wc -w for word count verification and convert (ImageMagick) or pdftoppm (poppler-utils) for PDF-to-image conversion. It also includes an inline Python script executed via python -c to facilitate visual review of generated files.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to dynamically install external software dependencies at runtime, specifically requesting the installation of the pdf2image library using the uv add command.
[INDIRECT_PROMPT_INJECTION]: The agent is directed to ingest untrusted data from the web and academic sources via the parallel-web and research-lookup skills. The skill lacks instructions for sanitizing this data before it is written into project files or processed, creating a surface for potential indirect prompt injection.
Ingestion points: External data retrieved through parallel-web and research-lookup skills (referenced in SKILL.md).
Boundary markers: None identified in the instruction set.
Capability inventory: File system access (write/edit), subprocess execution (shell commands), and tool invocation for external research.
Sanitization: No sanitization or validation logic is specified for the external content before interpolation into documents.

writing