The Agent Skills Directory

[COMMAND_EXECUTION]: The skill primarily functions by executing a local driver script (control-atlas.mjs) to automate UI tasks such as navigation, interaction, and state inspection within a desktop application.
[DYNAMIC_EXECUTION]: The skill includes an eval command that enables the execution of arbitrary JavaScript within the context of the driven application. This is explicitly documented for reading the DOM, clipboard, and other application attributes.
[DATA_EXFILTRATION]: The skill has the capability to access sensitive user data, including reading the system clipboard via JavaScript evaluation (eval "navigator.clipboard.readText()") and managing user profile data through the clone-userdata command.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from a live application environment (including chat transcripts, terminal output, and log files) and possesses powerful interaction capabilities (sending messages, UI manipulation, and code execution), creating an attack surface for indirect prompt injection.
Ingestion points: The skill reads application data via the snapshot (accessibility tree), eval (DOM and clipboard), and various log pane/terminal viewers described in the reference files.
Boundary markers: No specific delimiters or instructions to ignore embedded content were identified in the provided documentation.
Capability inventory: The toolset includes send (to transmit messages/prompts), click, type, eval (JS execution), and terminal interaction (shell-pane.md).
Sanitization: The documentation does not describe any sanitization or validation of the data ingested from the application environment.

verify-atlas