The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a local Python script scripts/check_package_consistency.py used to verify that quotes in response letters match the manuscript text. The agent is instructed to execute this script. Analysis of the script shows it uses only standard Python libraries (argparse, re, pathlib, etc.) and performs string matching operations without executing external shell commands or accessing the network.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to process untrusted data from external peer review reports and editor emails. This creates an attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the context via pasted emails and reviewer comments processed in SKILL.md and static/core/workflow.md.
Boundary markers: The skill uses explicit LaTeX macros (\ReviewerComment) and markdown headers (## Comment 1) to delimit external content.
Capability inventory: The skill can execute the provided consistency script and write files (marked manuscripts).
Sanitization: references/latex-templates.md provides explicit instructions to escape LaTeX-sensitive characters in user-provided content. The skill's strong enforcement of a professional, evidence-based tone further mitigates the risk of the model obeying instructions hidden within reviewer comments.
[DATA_EXPOSURE]: The skill accesses local manuscript files and author notes to perform its function. This is necessary for verifying quotes and drafting responses. There is no evidence of exfiltration; all network references in references/source-basis.md point to trusted, well-known domains (nature.com, springernature.com) for documentation purposes.

nature-response