The Agent Skills Directory

[COMMAND_EXECUTION]: The script uses git to identify modified UI files in the project. This is limited to local status checks and contains no user-controlled input.
[INDIRECT_PROMPT_INJECTION]: The skill processes UI code and may reflect matched fragments (e.g., non-standard spacing values) back to the agent in repair messages. The surface is highly restricted by regex patterns and the linter's specific purpose.
Ingestion points: scripts/ux_lint.mjs reads local project files via fs.readFileSync during hook execution.
Boundary markers: The output uses a rigid format (file:line [LEVEL] [ID] message) to clearly distinguish linter findings from other agent instructions.
Capability inventory: The script is limited to reading files, checking git status, and printing to standard output/error; it cannot perform network operations or write to files.
Sanitization: Findings are matched using specific, hardcoded regular expressions for CSS properties and Tailwind utility classes, preventing arbitrary code from being reflected.
[PERSISTENCE]: The skill instructs the agent to register hooks in .claude/settings.json, which causes the linting script to execute automatically on every file write or session stop. This is a standard and intended use of the platform's hook system to automate code quality checks.

ux-guardrails