The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation and logic for 'Auto Mode' (defined in SKILL.md, specs/auto-mode.md, and ccw-main.json) instruct the agent to 'skip clarification' and 'skip confirmation'. This effectively overrides standard agent safety protocols and human-in-the-loop checkpoints during automated workflows.\n- [COMMAND_EXECUTION]: The skill implements a 'Phase-Level Execution' mechanism that loads content from variable file paths (e.g., .claude/skills/*/phases/) and instructs the agent to execute these instructions 'inline'. Additionally, the skill parses user-supplied input to extract goals which are then passed as arguments to downstream skill calls, creating a surface for argument injection and automated execution of unintended commands.

ccw-chain