The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes external CLI utilities to manage source documents. \n
gemini.ts executes rclone to copy files from Google Drive to a local cache to avoid FUSE deadlocks during processing.\n
materialize-sources.ts executes the readwise CLI to search for and download article text for verification.\n
These executions are performed using spawnSync with argument arrays, which is a secure method for calling external processes that mitigates command injection risks.\n- [EXTERNAL_DOWNLOADS]: The skill downloads source materials as part of its core materialization process. \n
It fetches documents from Readwise and Google Drive using user-configured third-party CLI tools.\n
It uploads documents to Google's Gemini API (a trusted service) for the purpose of citation verification, which is the documented primary function of the skill.\n- [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes user-provided markdown drafts and bibliography files.\n
Ingestion points: Markdown drafts and BibTeX files (e.g., paperpile.bib) are read from the local file system.\n
Boundary markers: The LLM prompts are structured to request JSON responses, but there are no explicit delimiters around the interpolated citation claims to prevent instructions inside those claims from influencing the model.\n
Capability inventory: The skill can read local files, execute the rclone and readwise binaries, and communicate with the Google Gemini API.\n
Sanitization: The skill extracts and cleans citation text before sending it to the model to ensure the claim is clearly identified for the AI.

cite-check