The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests and processes untrusted data from Telegram messages. However, it implements significant mitigations.\n
Ingestion points: Incoming Telegram message text and file uploads are processed and passed to subagents in SKILL.md.\n
Boundary markers: The skill explicitly isolates user-provided content using <untrusted_telegram_question>, <untrusted_telegram_description>, and <untrusted_file_content> tags.\n
Capability inventory: The skill utilizes TaskCreate and TaskUpdate for agent coordination and Bash for network operations and script execution.\n
Sanitization: It includes instructions for agents to treat input as data only, performs HTML escaping for filenames, and enforces file extension allowlists and size limits.\n- [COMMAND_EXECUTION]: The skill uses the Bash tool to perform necessary operations for bot functionality.\n
It executes curl commands to download user-uploaded files from Telegram's official API servers.\n
It invokes a local Python script (markitdown-convert.py) to process downloaded files.\n- [EXTERNAL_DOWNLOADS]: The skill downloads content from well-known and official Telegram domains.\n
It fetches updates from api.telegram.org and downloads files from Telegram's file hosting service.\n- [DATA_EXFILTRATION]: The skill communicates with the Telegram API to send messages and deliver agent responses.\n
It correctly uses environment variables (e.g., TELEGRAM_BOT_TOKEN) for authentication, instructing users to manage these via a .env file to avoid hardcoding secrets.

telegram-polling