The Agent Skills Directory

Indirect Prompt Injection: The skill is designed to extract and execute setup, build, and test instructions directly from untrusted repository files like README.md, CONTRIBUTING.md, and package.json. This workflow creates a potential surface where malicious instructions embedded in a repository could influence agent behavior.
Ingestion points: Untrusted data enters the agent context via repository documentation and configuration files during the orientation and dependency installation phases.
Boundary markers: The instructions do not specify the use of delimiters or specific safety prompts to isolate extracted instructions from the agent's core guidelines.
Capability inventory: The skill possesses extensive capabilities, including shell command execution, system package installation (apt-get), and the ability to start backing services like databases.
Sanitization: The skill's primary objective is to "audit whether these documented instructions actually work," which involves running extracted commands verbatim without prior sanitization.
Command Execution: The skill performs administrative tasks such as installing system dependencies via apt-get, managing Node.js versions with nvm or fnm, and executing package manager commands. These actions are performed to facilitate the build process within a sandboxed environment.
External Downloads: The skill retrieves code via git clone and downloads third-party dependencies from public registries (NPM, PyPI, etc.). While these are standard development operations, they involve the retrieval of content from external sources.

build-repo