The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes instructions and mechanisms designed to override default agent behaviors and safety constraints.
Prompt templates in data/prompts/retro.md and data/prompts/create.md incorporate the #YOLO tag and explicit commands such as "Do NOT wait for user input" and "EVERYTHING SHOULD BE AUTOMATED."
The orchestrator's documentation in data/stop-hook-recovery.md directs sub-agents to "Never wait for user," bypassing the standard interaction model intended for human oversight.
[COMMAND_EXECUTION]: The skill makes extensive use of shell command execution and dynamic script generation.
The Python implementation uses subprocess.run to manage tmux sessions, git operations, and other system utilities.
It dynamically generates bash runner scripts and command files in the system's temporary directory (/tmp) and executes them via tmux respawn-pane (documented in src/story_automator/core/tmux_runtime.py).
It specifically configures command strings to include the claude --dangerously-skip-permissions flag for child sessions, which is explicitly intended to bypass standard permission prompts for tool usage.
[PROMPT_INJECTION]: The skill implements a mechanism to control the agent's lifecycle and prevent user intervention.
It installs a 'Stop hook' by modifying local configuration files such as .claude/settings.json or .codex/hooks.json (see src/story_automator/core/stop_hooks.py).
This hook executes a command that checks for an active automation marker and outputs a JSON block to the agent platform to prevent the user from stopping the session while the automation is active, representing a denial-of-service against user-initiated termination.

bmad-story-automator