The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions include standard pnpm commands (pnpm exec sanity blueprints deploy, pnpm exec sanity schema deploy) used to manage Sanity Studio infrastructure and schema deployments. These are routine development operations within the platform's ecosystem.
[PROMPT_INJECTION]: The skill manages an indirect prompt injection surface by design, as its primary purpose is to process untrusted source documents and terminology into prompts for translation.
Ingestion points: Untrusted data enters the context from l10n.glossary entries, l10n.style-guide instructions, and source document fields (e.g., articles, tags) within the Sanity dataset.
Boundary markers: The promptAssembly.ts logic implements structured sections for glossaries (Approved, DNT, Forbidden) to guide the LLM's behavior.
Capability inventory: The skill interacts with the Agent Actions Translate API and has the ability to write translated content back to the Sanity dataset using a provisioned robot token.
Sanitization: The architecture includes a specific utility, sanitizeTranslationValue.ts, designed to clean and validate AI output before it is committed to the database.
[DATA_EXFILTRATION]: The skill's evaluation framework resolves Sanity authentication tokens by checking the SANITY_AUTH_TOKEN environment variable and the standard Sanity CLI configuration directory (~/.config/sanity/). This is a legitimate requirement for authenticating API requests to the platform during development and testing.

sanity-l10n