The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied CSV and JSON files via bundled scripts to generate evaluation matrices and session scaffolds.
Ingestion points: Data enters via scripts/evaluate_matrix.py (CSV/JSON) and scripts/validate_register.py (JSON).
Boundary markers: Instructions in SKILL.md and references/responsible_ai.md emphasize that agent outputs are proposals, requiring mandatory human review and separation of processing stages.
Capability inventory: The local scripts are restricted to file system operations (writing JSON reports) and do not perform network requests or execute shell commands.
Sanitization: scripts/_common.py implements comprehensive validation, including a 5MB input size limit, text length restrictions, strict identifier regex matching, and protections against symlink attacks using O_NOFOLLOW and O_EXCL flags.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to fetch bibliographic metadata from a well-known academic repository to ensure citation accuracy.
Evidence: SKILL.md contains instructions to access https://arxiv.org/abs/2609.00065 or the arXiv API. This operation targets a well-known scientific service to verify author and version information for the skill's reference documentation.

scientific-brainstorming