The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The instructions in SKILL.md recommend fetching citation metadata from arXiv.org (a well-known repository for scientific papers) to ensure accuracy in references. This operation targets a well-known service for its intended primary purpose.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of local JSON and CSV files (e.g., hypothesis records and evidence ledgers).
Ingestion points: Data is read via read_json and read_csv_records in scripts/_common.py across all validation scripts.
Boundary markers: The scripts enforce strict schema validation, input size limits (2 MiB), and nesting depth limits (64 levels) to bound the execution context.
Capability inventory: The skill can read local files and generate Markdown scaffolds via scripts/generate_preregistration_scaffold.py.
Sanitization: The scaffold generator uses html.escape and Markdown character escaping to ensure that user-provided text remains inert in the generated output.
[SAFE]: The bundled Python scripts follow security best practices for local tool execution, including path resolution that rejects symlinks and URL-like schemes, atomic file writing with private permissions (0600), and the use of the standard library only to avoid external dependencies.

hypothesis-generation