The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acts as a document processing pipeline that ingests data from external files (PDF, Word, etc.) and interpolates this content directly into LLM prompts for translation and structural correction. Maliciously crafted content within an input document could attempt to override the system instructions or manipulate the translation process.
Ingestion points: scripts/preprocess_doc_with_docling.py and other scripts ingest untrusted data from user-provided file paths via the --input argument.
Boundary markers: The prompts in scripts/translate_ja.py and scripts/realign_doc_struct_with_llm.py do not use secure delimiters to isolate document content from system instructions.
Capability inventory: The skill can write files to the local system, execute local binaries via subprocess.run, and make outbound network requests to configured API endpoints.
Sanitization: No security-focused sanitization or filtering is performed on the ingested document text before it is sent to the LLM.
[COMMAND_EXECUTION]: The script scripts/convert_md_to_docx_with_docling.py uses subprocess.run to call the pandoc binary. While the arguments are passed as a list, which mitigates shell injection risks, executing external binaries with file paths derived from input parameters remains a relevant security surface.
[EXTERNAL_DOWNLOADS]: The skill makes outbound HTTP requests to services configured via DOCLING_SERVER_URL and OPENAI_BASE_URL. These services receive the full content of the documents being processed, necessitating that the user trust these configured endpoints with sensitive data.

translate-ja