The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions define workflows for ingesting untrusted repository artifacts into the agent context, which presents an indirect prompt injection surface.
Ingestion points: SKILL.md (Workflow 0) directs the agent to read README.md files, source code, and experimental outputs from the user's research repository.
Boundary markers: The workflow lacks explicit delimiters or instructions to ignore potential commands embedded within ingested project files.
Capability inventory: The skill uses shell tools (find, grep, pdflatex), python script execution, and network requests to academic APIs as documented across SKILL.md and templates/README.md.
Sanitization: No sanitization or filtering of repo-local content is described before interpolation into the prompt.
[COMMAND_EXECUTION]: The skill utilizes local shell commands for project exploration and LaTeX compilation as part of its primary functionality.
Evidence: SKILL.md (Workflow 0) contains command patterns using find, grep, and xargs to explore codebases.
Evidence: templates/README.md provides standard compilation commands including pdflatex, bibtex, and latexmk.
[EXTERNAL_DOWNLOADS]: The skill automates and references downloads from well-known academic and technology services.
Evidence: templates/neurips2025/Makefile includes an upgrade target that downloads style files from media.neurips.cc.
Evidence: SKILL.md suggests adding a search server from mcp.exa.ai via npx to facilitate real-time literature reviews.

ml-paper-writing