The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Instructions in SKILL.md, references/index.md, and references/quickstart.md recommend installing the application using curl -fsSL https://app.primeintellect.ai/prime-agent/install.sh | sh. Executing remote scripts by piping directly to a shell is a high-risk security practice.
[EXTERNAL_DOWNLOADS]: The documentation identifies several external resources used for software updates and configuration, including installation scripts from app.primeintellect.ai and release manifests from r2.dev (documented in references/settings.md).
[INDIRECT_PROMPT_INJECTION]: The skill documentation describes a system that automatically loads and follows instructions from AGENTS.md and CLAUDE.md files found in the current and parent directories (references/quickstart.md). This creates a vulnerability surface if the agent is used in a repository containing malicious instruction files.
Ingestion points: AGENTS.md and CLAUDE.md context files (specified in references/quickstart.md).
Boundary markers: Absent; the agent is instructed to load these files directly to guide its behavior.
Capability inventory: The system includes a persistent IPython kernel, arbitrary shell command execution via the ! prefix, and file modification tools (references/usage.md).
Sanitization: Not mentioned; there is no described mechanism for sanitizing or verifying the contents of workspace-level instruction files.

prime-agent-docs