The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest data from the current session—specifically API error messages, CLI tool outputs, and user statements—to draft feedback reports. If a failure was triggered by malicious external content (such as a poisoned API response or a malicious source file), the agent might inadvertently include malicious instructions or manipulated facts in the generated GitHub issue.\n
Ingestion points: The agent is instructed to process outputs from 'Crowdin skill instructions, CLI commands, MCP tool calls or API calls' and 'What the user said'.\n
Boundary markers: While the skill provides a structured Markdown template for the feedback, it lacks explicit delimiters or warnings for the agent to treat the ingested session data as untrusted text rather than instructions.\n
Capability inventory: The skill utilizes file system writes to ~/.crowdin/feedback/, command execution via the GitHub CLI, and generation of prefilled URLs to external websites.\n
Sanitization: The skill mandates strict privacy-focused sanitization (filtering tokens, project IDs, and source text), but it does not specify sanitization or escaping techniques to prevent instruction injection within the feedback body.\n- [COMMAND_EXECUTION]: The skill executes external command-line tools to perform its functions. It checks the authentication status of the GitHub CLI (gh auth status) and creates issues directly (gh issue create). Additionally, it may run crowdin --version to gather environment data. These tools are legitimate vendor-related or standard developer utilities, but their execution is triggered automatically based on session events or user requests.

feedback