The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages a local Node.js server using Bash scripts (start-server.sh, stop-server.sh). This includes spawning background processes and executing platform-specific browser launchers to facilitate the visual companion feature.
[INDIRECT_PROMPT_INJECTION]: The skill processes user interaction data from the visual companion by reading a local JSON-line file (events).
Ingestion points: The skill instructions in visual-companion.md direct the agent to read structured event data from $STATE_DIR/events created by the local server.
Boundary markers: The instructions lack explicit delimiters or warnings to ignore potential instructions embedded within the interaction data, relying on merging this data with terminal text.
Capability inventory: The agent can execute shell scripts for server lifecycle management, write HTML content for mockups, and read session state files.
Sanitization: While data is parsed as JSON by the server and read by the agent, no specific prompt-level sanitization or instruction filtering is performed on the ingested content.
[DYNAMIC_EXECUTION]: The skill dynamically generates HTML content fragments at runtime based on the brainstorming context. These fragments are written to a local directory and served by the companion server, which wraps them in a predefined frame template.

brainstorming