The Agent Skills Directory

[SAFE]: The skill serves as a front-end requirements engineering layer. It guides the agent to interview the user and extract project context (Known Knowns) from the existing environment before beginning work. This structured approach reduces errors and prevents the agent from making incorrect assumptions.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: In Phase 1, the skill scans local project metadata such as CLAUDE.md and handoff.md to establish current project status. This data is used solely to populate the local specification card and facilitate the interview process. No external network exfiltration or unauthorized data access was found.\n- [PROMPT_INJECTION]: The skill includes instructions to interpret voice-input homophones (e.g., '阿滴Q' for 'RDQ') and to prompt the user to start an RDQ session for large tasks. These instructions are functional and designed to enhance the agent's usability rather than bypass safety guardrails.\n- [INDIRECT_PROMPT_INJECTION]: The skill reads external data (project files and user input) which could theoretically contain injection attempts. However, the skill implements a robust mitigation strategy: the 'status: draft' gate. The agent is strictly instructed not to execute any task until the user has reviewed and manually confirmed the generated specification card, ensuring a human-in-the-loop security checkpoint.\n- [COMMAND_EXECUTION]: The skill performs standard file system operations to store its specifications (e.g., creating the rdq/ directory and markdown files). It does not execute arbitrary shell commands or use dynamic execution patterns that could be exploited.

rdq