The Agent Skills Directory

External Dependency Management: The skill installs Node.js packages using npm install. This is a standard development task, and the skill implements significant safeguards: it forces the use of the official npm registry, utilizes exact versions, and performs post-installation lockfile and provenance verification to ensure the artifact matches the previously validated state.
Remote Content Retrieval: The skill can fetch package documentation from HTTPS URLs to verify framework compatibility. To mitigate risks, the process is designed to reject non-HTTPS or local hosts, limit the payload size, and strip active HTML content. The resulting text is treated strictly as untrusted evidence, with explicit instructions for the agent to ignore any embedded directives.
Indirect Prompt Injection Surface: The skill ingests data from project configuration files and external documentation content, which could potentially contain malicious instructions.
Ingestion points: Project files such as powerpages.config.json and external URLs provided for package evidence (e.g., modeEvidenceUrl).
Boundary markers: The skill explicitly instructs the agent to treat these inputs as "untrustedProjectData" or "untrusted evidence only" and to disregard any instructions or tool calls found within those values.
Capability inventory: The agent has the ability to execute shell commands, install npm packages, navigate via Playwright, and write to the project filesystem.
Sanitization: The skill uses dedicated validation scripts and strips active content from remote fetches during the evidence collection phase.
Automated UI Verification: It uses Playwright to perform visual and functional audits of the localized site on a development server. This involves browser navigation and interaction, which is a common practice for UI testing and is performed within a controlled transaction-aware environment.
User Authorization Gates: Critical actions, such as package installation, implementation plan approval, and enabling newly added locales, are protected by explicit AskUserQuestion gates, ensuring the human developer remains in control of project modifications.

add-localization