The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from EffectCraft projects, such as layer names, property paths, and expressions, which are ingested through methods like layer.tree, prop.get, and ui.inspect (referenced in SKILL.md and references/control-protocol.md).
Ingestion points: Project summary, composition info, layer tree, and property values.
Boundary markers: No explicit delimiters or instructions are provided to the agent to disregard potential commands embedded within this project-level data.
Capability inventory: The skill possesses powerful capabilities including run_script, engine.execute, and ui.menu.invoke which could be triggered or influenced by malicious data injected into project properties.
Sanitization: No sanitization or validation of the ingested strings is mentioned before they are processed by the agent's logic.
[DYNAMIC_EXECUTION]: The skill utilizes the run_script command and the engine.execute method to perform programmatic actions within the host application. While necessary for the skill's purpose (visual effects automation), this allows for the execution of dynamically generated logic which can be misused if the agent is misled by project data.
[COMMAND_EXECUTION]: The instructions in SKILL.md and references/control-protocol.md guide the agent to use shell-based tools such as effectcraft-cli, printf, and nc (netcat) to manage sessions, start the application with specific flags, and communicate with the local TCP port on 127.0.0.1.

effectcraft