The Agent Skills Directory

[COMMAND_EXECUTION]: The guide suggests running claude --dangerously-skip-permissions. This flag is a feature of the Claude Code CLI that suppresses interactive permission prompts for file and network access during the session.
[EXTERNAL_DOWNLOADS]: The skill requires downloading the telegram@claude-plugins/official plugin to enable communication with the Telegram API.
[PROMPT_INJECTION]: The integration creates an indirect prompt injection surface where external data enters the agent context. (1) Ingestion point: Telegram bot messages (SKILL.md). (2) Boundary markers: No delimiters are specified in the prompts. (3) Capability inventory: The agent can execute commands and write files as demonstrated in the test phase (Step 5-2). (4) Sanitization: The skill addresses this by instructing the user to enable a mandatory 'allowlist' access policy to restrict inputs to known accounts.

claudeclaw