The Agent Skills Directory

[SAFE]: The skill implements a robust security posture using the 'protect-pass' mechanism, which replaces sensitive structural elements (URLs, HTML, placeholders) with opaque sentinels (e.g., ⟦P:NN⟧) before processing. This prevents the model from inadvertently modifying or leaking structural code during translation.
[DATA_EXPOSURE]: The skill accesses local file paths for project glossaries (e.g., docs/i18n/glossary-{target_locale}.md). This is a legitimate and scoped file access pattern required for its primary function. No evidence of credential harvesting or unauthorized file access was found.
[PROMPT_INJECTION]: The skill uses clear XML-like boundary markers (<TRANSLATE_THIS>) to isolate untrusted source text from operational instructions. This is a recognized mitigation strategy for indirect prompt injection, ensuring the model treats the input as data to be translated rather than instructions to be executed.
[EXTERNAL_DOWNLOADS]: The skill documentation mentions optional web search capabilities for terminology resolution (L3 tier). These are standard features for professional translation workflows and target well-known search services. No unauthorized or suspicious remote code downloads are present.
[COMMAND_EXECUTION]: Analysis of SKILL.md and all referenced protocols confirms the absence of dynamic context injection (!command syntax) or any subprocess execution patterns. The skill relies entirely on role-based LLM prompts.

translation-creative