The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external public URLs and local repository files to generate documentation drafts.
Ingestion points: Processes data from public HTTPS documentation URLs, local repository manifests, scripts, and documentation files as described in SKILL.md and references/source-workflow.md.
Boundary markers: Employs independent sub-agents to isolate research tasks. Instructions in references/deriving-context.md explicitly mandate treating source content as evidence, not as instructions to execute or expand scope.
Capability inventory: The skill possesses capabilities to read local files, execute read-only CLI commands, call MCP tools, and write files to the repository.
Sanitization: Implements a 'CLI access contract' in references/source-workflow.md that uses a runtime allowlist and subprocess argument vectors to prevent command injection from retrieved data.
[COMMAND_EXECUTION]: The skill facilitates the execution of design-system CLI tools and package managers.
Evidence: references/source-workflow.md outlines the 'Design-system CLI access' contract, allowing execution of user-named, already installed executables.
Security controls: All invocations are required to use a non-shell subprocess interface with shell evaluation disabled. It explicitly rejects shell metacharacters and expansions.
[EXTERNAL_DOWNLOADS]: The skill can download and install packages and browser-related tooling.
Evidence: references/browser-tooling.md details the process for installing user-named packages and browser drivers like Playwright.
Security controls: Every installation requires explicit user approval, providing the exact package, version, and command before execution. It enforces the use of the repository's existing package manager.

setup-skills