The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves reading untrusted content from research repositories, which creates a surface for indirect prompt injection attacks.
Ingestion points: The "Project Understanding" phase in SKILL.md (Workflow 0) directs the agent to explore the workspace using read_file and search_project to understand claims and results.
Boundary markers: Absent. The instructions do not include explicit delimiters or warnings to ignore instructions found within the project's files.
Capability inventory: The skill possesses file system read/write and search capabilities (read_file, write_file, search_project).
Sanitization: Absent. Data from ingested files is used directly in the prompt context to generate drafts and revisions.
[EXTERNAL_DOWNLOADS]: The skill references downloads from official conference sources for configuration and formatting tasks.
templates/neurips2025/Makefile uses curl to fetch official conference style files from media.neurips.cc.
SKILL.md provides guidance for configuring an external academic search tool via https://mcp.exa.ai/mcp.

ml-paper-writing