The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external and local academic documents, such as PDFs, Markdown files, and LaTeX manuscripts, to extract structural and rhetorical patterns. These documents represent an ingestion point for potentially malicious instructions intended to influence the agent's synthesis or research output.
Ingestion points: Local file indexing performed by scripts/reference_inventory.py and manuscript reading by scripts/humanize_check.py and the sub-agent task logic.
Boundary markers: The skill employs task-specific sub-agents (Scene Analyst, Exemplar Learner, SOTA Mapper) and limits the context passed to each, which provides some degree of isolation.
Capability inventory: The skill uses file-read operations, sub-agent tool calls, and web search capabilities based on user-provided URLs.
Sanitization: Although scripts/_paper_spine_utils.py includes functions to normalize LaTeX and Markdown structure, no security-specific sanitization or filtering of semantic content for prompt injection is implemented.
[COMMAND_EXECUTION]: The skill relies on the execution of local Python scripts to automate research tasks.
Evidence: The skill invokes scripts/reference_inventory.py to index local materials and scripts/humanize_check.py to validate manuscripts. The instructions also mention scripts/style_metrics.py, which is referenced but not provided in the source files. These scripts exclusively use Python standard library modules (e.g., re, argparse, pathlib, json).

paper-spine-research