The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided descriptions to generate plugin manifests and capability logic, creating a surface for indirect instructions to influence the generated output.
Ingestion points: User requests specifying plugin IDs, names, capability details, and logic are processed within the skill to generate artifacts.
Boundary markers: The skill contains explicit instructions to never install, enable, trust, or execute the generated plugin or its dependencies, leaving installation to the user as a manual step.
Capability inventory: The skill utilizes write_file and edit_file for creating workspace artifacts and shell execution for plugin validation via the muse CLI.
Sanitization: Strict validation rules are applied to plugin IDs using a specific ASCII grammar and regex, and artifact paths are limited to relative UTF-8 paths within the plugin root, rejecting absolute paths and parent traversal.
[COMMAND_EXECUTION]: The skill executes the muse CLI tool (muse skills validate and muse plugins validate) to verify the generated plugin structure. These commands are executed with structured arguments and operate on paths that have been validated to prevent traversal or escaping.
[DYNAMIC_EXECUTION]: The skill generates executable scripts, such as shell hooks (hooks/pre-check.sh) and MCP servers (mcp/server.py), based on provided examples and user requirements. While the skill is strictly forbidden from executing these generated files, their creation represents a surface for dynamic code generation from templates.

create-plugin