The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with the local system by executing the claude CLI binary. It assembles command-line arguments based on user input, including potentially high-risk flags like --dangerously-skip-permissions and --allow-dangerously-skip-permissions which bypass security prompts.
[EXTERNAL_DOWNLOADS]: The references/install.md file instructs the agent to download and run installation scripts from https://claude.ai/install.sh (macOS/Linux/WSL) and https://claude.ai/install.ps1 (Windows). These URLs originate from an established and recognized service provider.
[INDIRECT_PROMPT_INJECTION]: The skill interpolates user-provided prompts into shell commands, creating a surface for injection if the quoting is bypassed or if the shell handles special characters unexpectedly.
Ingestion points: User-supplied <prompt> and <follow-up> variables are used in SKILL.md and references/sessions.md.
Boundary markers: The instructions specify that the agent must "quote the prompt literally" to maintain integrity.
Capability inventory: The skill performs subprocess execution of the claude binary and potentially other management commands.
Sanitization: Relies on the agent correctly quoting the input string to prevent shell metacharacter execution.

claude-code