The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The architecture establishes a pipeline for ingesting and processing translation data from both local files and remote services.\n
Ingestion points: The skill references reading translation files (e.g., en.json) in scripts like generate-types.ts and fetching runtime data via the Lokalise Over-the-Air (OTA) CDN.\n
Boundary markers: The provided examples do not include specific delimiters or instructions to the agent to disregard potential prompt injection attempts embedded within translation strings.\n
Capability inventory: While the skill's allowed tools are limited to Read and Grep, the recommended implementation includes scripts that perform file system writes and network requests, creating a surface where malicious strings could influence downstream processes.\n
Sanitization: The implementation guide does not currently feature validation or sanitization logic to verify the safety of translation values before they are used for code generation or application rendering.\n- [EXTERNAL_DOWNLOADS]: The skill incorporates functionality to download translation assets from Lokalise, a well-known translation management service.\n
The implementation correctly targets official domains such as ota.lokalise.com for fetching updates.\n
These operations are documented neutrally as part of the intended functionality for a localization workflow.

lokalise-reference-architecture