The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill is designed to clone and execute external software projects to generate showcase materials. This process involves running package installation commands (e.g., npm install, pnpm install) and starting development servers or Docker containers from the cloned source code. Although safety protocols are detailed, such as using throwaway clones and ensuring local-only database connections, executing arbitrary code from untrusted repositories is inherently risky.
[EXTERNAL_DOWNLOADS]: The skill uses git clone to download external project code from repository URLs provided by the user. This facilitates the retrieval of untrusted content into the agent's operating environment.
[COMMAND_EXECUTION]: The skill uses various shell commands to manage project lifecycles, including git, docker, docker compose, and Node.js package managers. It also executes internal utility scripts like scripts/screenshot-session.mjs to coordinate these operations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes untrusted codebases, which creates an attack surface for indirect prompt injection. Malicious content within the cloned repositories could potentially influence the behavior of analysis agents.
Ingestion points: Project files are cloned from external repositories into ~/.lt-showroom/<slug>/checkout.
Boundary markers: The skill does not explicitly mention using delimiters or safety instructions when passing repository content to the analysis agent.
Capability inventory: The environment has access to tools including node, docker, bash, and network access via git and curl.
Sanitization: While DOM anonymization is used for visual captures, the skill does not describe sanitization of the repository source code before analysis by the LLM.

creating-showcases