The Agent Skills Directory

[SAFE]: The skill implements a robust architecture for translating markdown documents while strictly preserving technical structure. It uses an AST (Abstract Syntax Tree) parsing approach to isolate prose from structural elements.
[SAFE]: Structural integrity is protected via a 'protect-pass' mechanism that replaces sensitive spans (code, math, URLs, HTML tags) with opaque sentinel tokens (e.g., ⟦P:NN⟧). This prevents the LLM from inadvertently modifying or leaking data from these blocks during the translation process.
[SAFE]: The skill uses established, well-known resources for its glossary (e.g., Pontoon, GNOME, NAER). It supports an optional web search for term resolution (L3 tier) which is documented as using standard platform search tools.
[SAFE]: Prompt security is enhanced by role-based decomposition (Writer, Critic, Reviser) and the use of explicit boundary markers (<TRANSLATE_THIS>) to isolate the active translation chunk, significantly reducing the surface for indirect prompt injection from untrusted document content.
[SAFE]: Verification gates (M1-M3, S1-S2) provide a multi-layered security and quality check, including deterministic regex checks for placeholder parity and deterministic linters for residual source language text.

translation-doc