The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill exposes a run_script tool that enables the execution of arbitrary JavaScript code within the host engine's scripting environment (e.g., app, project, layers manipulation).
[COMMAND_EXECUTION]: The command registry includes file.executeFile, which allows the agent to execute arbitrary files on the local system.
[EXTERNAL_DOWNLOADS]: The skill's setup process (bootstrap.py and desktop.py install) involves downloading and installing binary executables and CLI tools from remote repositories.
[INDIRECT_PROMPT_INJECTION]: The skill processes structured plan files (craft-command-plan/v1) and project data that can be provided by external sources. These files serve as ingestion points for instructions that, when combined with high-impact capabilities like script execution and file writing, create an attack surface for indirect injection attacks.
Ingestion points: Plan files (e.g., PLAN.json), project files (.ecproj), and imported media assets.
Boundary markers: The skill relies on a structured JSON schema for command plans, though it lacks explicit prompt-level delimiters to segregate untrusted data.
Capability inventory: Extensive capabilities including run_script (JS execution), file.executeFile, and various file system write operations (save_project, render_frame).
Sanitization: A check command is provided to validate the structure and membership of plans, but it does not sanitize the logic within scripts or command parameters.

effectcraft-cli-composition