The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs several shell operations to manage its environment and process data.
It installs the gdown package using pip install gdown in references/monolingual-sources.md.
It executes a local Node.js script (yomitan.mjs) located in a sibling skill directory to unpack, get, and add dictionary entries.
It uses directory management commands like mkdir -p and cd to set up a local cache for downloaded files.
[EXTERNAL_DOWNLOADS]: The skill retrieves dictionary archive files (.zip) from external, non-vendor locations.
It fetches content from Google Drive folders (1tTdLppnqMfVC5otPlX_cs4ixlIgjv_lH and 1LXMIOoaWASIntlx1w08njNU005lS5lez) using the gdown utility.
It references the MarvNC/yomitan-dictionaries GitHub repository as a source for direct download links.
While Google Drive and GitHub are well-known services, the specific content is hosted by community members and represents third-party data dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to read and process external dictionary definitions, which creates an attack surface for indirect prompt injection.
Ingestion points: The agent reads definition text extracted from third-party ZIP files via the node yomitan.mjs get command.
Boundary markers: The instructions do not specify strict boundary markers or sanitization for the definitions being read.
Capability inventory: The agent has the ability to execute Node.js scripts, write to the filesystem (via yomitan.mjs add/pack), and perform network operations via gdown.
Sanitization: There is no automated sanitization of the dictionary content. However, the skill implements a manual 'Quality Gate' requiring the agent to verify that target-language senses trace to quoted monolingual senses, which serves as a human-in-the-loop mitigation.

japanese-gloss-craft