The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill's setup instructions direct users to download and install the Model Context Protocol (MCP) server and a Roblox Studio plugin from the vendor's repository (ShiroKSH/skills) using git and npx.
[REMOTE_CODE_EXECUTION]: The skill provides tools for executing Luau code within a connected Roblox Studio session. Security is managed through instructions that mandate the agent explain any code and risks to the user first, and specifically requires a confirmToken ('RUN_STUDIO_TEST') for executing functional tests.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external Roblox Studio instances, creating a surface where malicious instructions could be embedded in property values or scripts.
Ingestion points: Hierarchy and property data are read from the active Studio session via tools like studio_get_tree and studio_get_properties as described in SKILL.md.
Boundary markers: No specific delimiters or XML tags are explicitly defined for separating ingested instance data from the prompt context.
Capability inventory: The skill possesses extensive capabilities including setting properties, executing code in Studio, and writing to the local filesystem via Rojo.
Sanitization: The instructions require the agent to review every script source and dry-run every plan before application to mitigate the risk of untrusted data influencing actions.
[DYNAMIC_EXECUTION]: Functional tests are executed using the studio_run_test tool. The documentation notes that these tests execute ordinary project code rather than utilizing a sandbox, which is a necessary behavior for testing Roblox environments. This risk is addressed by instructions to use isolated working copies and mock persistent writes.
[COMMAND_EXECUTION]: The documentation includes shell commands for the initial installation of the skill's environment and dependencies, including the use of npm for building the local bridge plugin.

roblox-studio