The Agent Skills Directory

[SAFE]: The skill is a well-structured instructional framework for technical planning. It relies on standard platform features (like Claude Code's Plan mode) and includes explicit completion criteria and failure mode handling, demonstrating a secure-by-design approach.\n- [DATA_EXFILTRATION]: The instructions permit the agent to inspect repository environment files (e.g., .env) and local configuration directories to gather project context. This is mitigated by explicit safety rules that strictly prohibit the inclusion of secrets, credentials, or private identifiers in the generated implementation specs or example code.\n- [PROMPT_INJECTION]: As a planning tool that ingests external repository data (such as README files and issue descriptions), the skill presents an indirect prompt injection surface. An attacker could attempt to place malicious instructions in project documentation to influence the generated implementation plan. The skill addresses this by enforcing structured markdown templates, mandatory user verification checkpoints, and explicit 'source challenge' steps that require cross-referencing instructions against repository reality.\n- [COMMAND_EXECUTION]: The skill generates validation commands and implementation plans involving shell execution. These are presented to the user as testable criteria and are intended for review before execution, aligning with the principle of human-in-the-loop verification.

claude-spec-interviewer