The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill implements a bootstrapping mechanism in scripts/phoenix-bootstrap.cjs that downloads platform-specific native engine binaries (phoenix-ui) from the author's infrastructure if they are missing from the local cache (~/.phoenix-ui/bin). The download process enforces HTTPS and performs mandatory integrity verification using SHA-256 digests and file size checks against a hardcoded manifest in scripts/ENGINE.json.
[COMMAND_EXECUTION]: Spawns the verified native engine binary using the Node.js child_process module to perform design detection, auditing, and image processing. This is the skill's primary functional path and operates within the user's project context to provide localized design services.
[PERSISTENCE]: The skill can install design detector hooks into various editor configuration files (e.g., .claude/settings.local.json, .cursor/hooks.json, .github/hooks/phoenix-ui.json) to automate design linting during development. These hooks are user-controllable, managed via the $phoenix-ui hooks command, and transparently documented in reference/hooks.md.
[DYNAMIC_EXECUTION]: The skill supports an interactive 'live mode' that injects a helper script (live.js) into project HTML files via a local server to facilitate real-time UI manipulation and variant swaps in the browser. This is standard functionality for advanced frontend design tools and is restricted to the local development environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests project source code and rendered data to generate design critiques and audits.
Ingestion points: Project source files scanned by phoenix-ui detect (e.g., as described in reference/critique.md).
Boundary markers: Instructions in reference/critique.md mandate isolating LLM-based design reviews from deterministic detector findings to prevent anchoring.
Capability inventory: Binary execution, filesystem writes (DESIGN.md), and browser script injection.
Sanitization: The tool uses structured data (JSON) for scanner outputs and utilizes modular sub-agents with limited, task-specific scopes to process external content safely.

phoenix-ui