The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for training tokenizers on external datasets, which introduces a surface for indirect prompt injection.
Ingestion points: The skill ingests raw text through file paths (e.g., data.txt) or Python iterators passed to the training functions (SKILL.md, references/training.md).
Boundary markers: There are no specific instructions or delimiters used to separate the training data from the agent's operating instructions.
Capability inventory: The skill utilizes the sentencepiece and transformers libraries to read training corpora and write model and vocabulary files to the local file system.
Sanitization: The instructions do not include steps for sanitizing or filtering the content of the training data before processing.
[PRIVILEGE_ESCALATION]: The manual C++ installation section includes commands to install the compiled library system-wide using elevated privileges.
Evidence: The use of sudo make install in the build workflow in SKILL.md.
[EXTERNAL_DOWNLOADS]: The skill fetches the project source code from Google's official GitHub repository during the manual installation process.
Evidence: git clone https://github.com/google/sentencepiece.git in SKILL.md.

sentencepiece