The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's Dockerfile templates and project initialization scripts fetch the Claude CLI tool from the official Anthropic domain at https://claude.ai/install.sh. This is a trusted and well-known installation source for the skill's core dependency.
[COMMAND_EXECUTION]: The skill's primary function is to operate an autonomous loop that executes arbitrary shell commands via the Claude CLI to implement features and tests. This execution is protected by a mandatory sandbox verification script (ralph.sh) that checks for containerization (Docker), network isolation, and non-root user privileges. The skill will refuse to start if these safety criteria are not met.
[DATA_EXFILTRATION]: The sandbox enforcement module proactively checks for network isolation and attempts to confirm that sensitive host directories—such as .ssh, .aws, and .config/gcloud—are inaccessible to the autonomous agent, mitigating the risk of credential theft or data exfiltration.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it autonomously processes untrusted data from project specifications and plan files. The mandatory hardware and software sandbox environment serves as the primary defense against such threats. \n
Ingestion points: IMPLEMENTATION_PLAN.md, specs/*.md, and project source files.\n
Boundary markers: Absent.\n
Capability inventory: Arbitrary shell execution, file system manipulation, and git operations via Claude CLI.\n
Sanitization: Absent.

ralph-wiggum