The Agent Skills Directory

[PROMPT_INJECTION]: The skill has a surface for indirect prompt injection because it instructs the agent to ingest and act upon data from external files that could be modified by third parties.
Ingestion points: The agent is directed to read docs/BOARD.html for task prioritization, tests/scenarios/*.md for executing scripted playtests, and console output blocks for analyzing telemetry data.
Boundary markers: The instructions do not specify the use of delimiters or boundary markers to separate data from instructions when reading these external files.
Capability inventory: The agent has the ability to execute shell commands (grep, find, ls), run internal testing tools (/roblox-testing:test), and manage the project toolchain (aftman install).
Sanitization: There are no explicit instructions for the agent to sanitize or validate the content of these external files before using them to drive logic.
[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of external development tools to enable headless simulation capabilities.
Pattern: The guidance includes adding lune = "lune-org/lune@<version>" to aftman.toml and executing aftman install to retrieve the binary.
Source: The download targets the lune-org organization on GitHub, which is a well-known provider of Luau runtime environments for the Roblox ecosystem.
Guardrail: The skill explicitly notes that adding new tools is an owner-gated action requiring approval, providing a human-in-the-loop safety check.
[COMMAND_EXECUTION]: The skill utilizes standard shell commands to perform project reconnaissance and manage local development environments.
Evidence: Includes instructions to use grep, find, and ls for locating specific modules or files, rojo serve to start a local development server, and aftman install for toolchain management.
Context: These commands are used within the scope of intended developer workflows and are restricted to the local project environment.

roblox-research-frontier