The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied design documents and visual artifacts, creating a surface for potential indirect prompt injection attacks.
Ingestion points: Ingests DESIGN.md (Markdown/YAML) and external visual artifacts (mockups, screenshots) from the repository worktree.
Boundary markers: The instructions mandate strict adherence to a routing map, an entry gate for source grounding, and a comprehensive evaluation rubric with binary readiness checklists.
Capability inventory: The skill executes shell commands via subprocess.run (Python) and npx (Node.js), and performs file writes to local directories (.codex/artifacts/).
Sanitization: Validation is performed using an official linter (@google/design.md) and a deterministic Python script to enforce structure, required sections, and token binding.
[DYNAMIC_EXECUTION]: The validation utility generates and executes a temporary Node.js script to interface with the design linter at runtime.
Evidence: scripts/validate-design-md-completeness.py writes a hardcoded JavaScript wrapper (parse.mjs) to a temporary directory and executes it using node to parse linter output.
[EXTERNAL_DOWNLOADS]: The skill fetches a specific version of a design linter from the official npm registry during validation.
Evidence: scripts/validate-design-md-completeness.py executes npm install @google/design.md@0.4.0. This download targets a well-known package from a trusted vendor.
[COMMAND_EXECUTION]: Instructions guide the agent to execute shell commands for linting, diffing, and validating UI contracts.
Evidence: references/design-md-contract.md and references/multi-surface.md provide explicit command forms for npx, designmd lint, and python3 scripts/validate-design-md-completeness.py to gate specification readiness.

ui-spec