The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation scripts (scripts/install.sh and scripts/install.ps1) download and execute the uv toolchain installer from a well-known service (astral.sh) using shell piping techniques (curl | sh and irm | iex).
[EXTERNAL_DOWNLOADS]: The skill downloads its core application and canvas components from the vendor's official GitHub repository (ReScienceLab/super-prototyping) during the installation process and when starting the canvas in headless mode.
[PERSISTENCE]: The installation process modifies shell profile configuration files (such as ~/.zshrc or ~/.bash_profile) to add the tool's installation directory to the system PATH. Additionally, it creates symbolic links for the skill in various agent-specific directories (e.g., ~/.claude/skills) to ensure the skill remains available across sessions.
[DYNAMIC_EXECUTION]: The skill utilizes a convention where board folders contain a gen.py script. The agent is instructed to edit these Python scripts and execute them locally to regenerate artboard content, representing a form of dynamic execution of local code.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, including HTML files, tldraw canvas records, and annotated screenshots provided by users. This creates an attack surface for indirect prompt injection. To mitigate this, the skill provides specific instructions to render artboards within sandboxed iframes (sandbox="") and to avoid sensitive flags like allow-same-origin.
[COMMAND_EXECUTION]: The skill makes extensive use of the sp command-line utility and other system tools (including python3, bun, curl, and hdiutil) to interact with the local file system and manage the canvas application environment.

sp-canvas