The Agent Skills Directory

[SAFE]: The skill implements a robust 'Human-in-the-Loop' philosophy, requiring explicit user approval at multiple stages (Scope, Analysis, Literature, Learning, Draft) before proceeding.
[PROMPT_INJECTION]: The skill contains a dedicated sanitization module in utils.py and compliance.py specifically designed to detect and redact prompt injection patterns (e.g., 'ignore previous instructions', 'you are now') and hidden text techniques (zero-width characters, LaTeX white-text hiding).
[COMMAND_EXECUTION]: Orchestrates various local skills (assess, dogpile, arxiv, memory, fixture-graph) using subprocess.run. Command arguments are constructed from validated internal configurations rather than raw user input, mitigating risk of command injection.
[EXTERNAL_DOWNLOADS]: Fetches academic metadata and paper information from well-known services including the arXiv API, CrossRef API, and Semantic Scholar. These operations are performed via httpx with neutral data-fetching intent.
[SAFE]: Includes a 'Quality Dashboard' and verification tools to detect hallucinated references or ungrounded claims, ensuring the output aligns with the analyzed codebase.

create-paper