The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides instructions to download and execute installation scripts for various development tools. Evidence includes installers from official domains such as claude.ai and chatgpt.com in examples/24gb-nvidia-local-plus-cloud.md and references/ai-coding-agents.md. It also references installation scripts for Ollama from ollama.com in SKILL.md and the Zed editor from zed.dev in references/ai-coding-agents.md.
[EXTERNAL_DOWNLOADS]: Fetches a variety of AI-related packages and container images from public registries. This includes Python packages like torch, vllm, and sglang from PyPI (references/ml-frameworks.md), Node.js packages for coding agents and MCP servers from npm (references/ai-coding-agents.md), and container images for TabbyAPI and Ollama from GitHub Container Registry (templates/gpu-compose.yaml).
[COMMAND_EXECUTION]: Utilizes local shell commands for system configuration and hardware management. Examples include hardware discovery commands like nvidia-smi and rocm-smi in SKILL.md, system service management using systemctl for local LLM engines in references/local-llm-inference.md, and environment setup using tools like uv, pixi, and docker.
[INDIRECT_PROMPT_INJECTION]: The skill identifies the potential for indirect prompt injection and provides detailed mitigation strategies.
Ingestion points: Data enters the agent context through MCP servers (github, fetch, filesystem) that read external content (references/mcp-servers.md).
Boundary markers: The skill recommends configuring safety defaults in settings files like ~/.claude/settings.json to define explicit permission boundaries.
Capability inventory: The environment utilizes tools with capabilities for subprocess execution (npx, curl, pip), file system writes, and network access.
Sanitization: It advocates for the use of OS-level sandboxing (landlock/seccomp), least-privilege scoping, and manual approval for sensitive commands.

linux-ai-dev-stack