The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md define a clear 'Reviewer-minded Editor' role with specific constraints regarding data integrity (not inventing numbers, preserving citation keys). There are no attempts to bypass safety filters or extract system prompts.
[DATA_EXFILTRATION]: The skill operates entirely on local workspace files such as sections/*.md and configuration JSONL files. The Python script scripts/run.py does not contain any network-related code (no requests, urllib, or socket operations) and follows a strictly local read/write cycle.
[REMOTE_CODE_EXECUTION]: No remote script downloads or piped shell executions were found. The Python logic uses standard libraries and regex for text processing.
[COMMAND_EXECUTION]: The script uses standard file system operations via pathlib and does not spawn subshells or execute arbitrary system commands.
[CREDENTIALS_UNSAFE]: No hardcoded API keys, tokens, or sensitive configuration paths (like .env or .ssh) are accessed or required by the skill.
[SAFE]: All external resources and logic reside within the skill's own package or the local repository environment. The behavior matches the stated purpose of improving numeric claim hygiene in technical surveys.

evaluation-anchor-checker