The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the agent with the ability to execute arbitrary shell commands and send keystrokes to other terminal panes and tabs using the prowl send and prowl key commands. This allows one agent to control the environment and actions of sibling agents.
[DATA_EXFILTRATION]: The prowl read and prowl agents read commands enable an agent to scrape terminal output and screen content from other agent sessions. This can lead to the exposure of sensitive data, environment variables, or secrets that another agent may have printed to its terminal history.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates inter-agent communication where one agent can read the output of another.
Ingestion points: Terminal output and agent snapshots are ingested via prowl read and prowl agents read (SKILL.md).
Boundary markers: No specific delimiters or boundary warnings are established for separating untrusted terminal content from instruction context.
Capability inventory: Full shell command execution via prowl send, keystroke injection via prowl key, and remote agent tasking via prowl agents dispatch are available across all scripts.
Sanitization: No sanitization or escaping mechanisms are described for processing the content retrieved from other agent panes.
[PERSISTENCE]: The skill describes the use of prowl skills install, which creates persistent links for the skill in multiple agent-specific directories (e.g., ~/.claude/skills, ~/.codex/skills, ~/.trae/skills). This ensures the skill's instructions remain available across different agent platforms and application updates.
[DYNAMIC_EXECUTION]: The prowl agents dispatch command allows for the dynamic delivery of new prompts and instructions to agents in other panes at runtime, enabling complex multi-agent workflows but also allowing for the execution of dynamically generated instructions.

prowl-cli