The Agent Skills Directory

[SAFE]: The skill is a standard utility for Japanese text correction. It operates by reading local files, processing the text through the Google Gemini API, and writing an annotated report to a local directory.
[SAFE]: Data handling is localized and transparent. The script sends the input text to a trusted external service (Google Gemini API) for its core functionality and does not attempt to access sensitive system directories or credentials outside of those required for API authentication.
[SAFE]: The script uses yaml.safe_load for parsing configuration files, which protects against YAML-based code execution vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface as it ingests content from external files provided by the user.
Ingestion points: scripts/proofreading_agent.py (line 331) reads content from the file system via the --input parameter.
Boundary markers: Absent. The input text is numbered but not enclosed in XML tags or other delimiters to prevent the LLM from interpreting parts of the input as instructions.
Capability inventory: The agent can read and write local files and communicate with the Gemini API. It does not have arbitrary code execution capabilities.
Sanitization: No sanitization or filtering is performed on the ingested text before interpolation into the prompt.

proofreading-agent