The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied requirements and prototype images to generate application code and deployment scripts, creating a surface for indirect prompt injection if the inputs contain malicious instructions targeting the code generation logic.
Ingestion points: The skill reads /workspace/prd.md and UI prototype images from /workspace/prototypes/ as the primary source for all generated artifacts.
Boundary markers: There are no explicit instructions or delimiters used when processing these files to differentiate between data and instructions.
Capability inventory: The skill possesses extensive capabilities including file system write access, execution of generated shell scripts (start.sh), and network operations via Python scripts.
Sanitization: The skill does not implement sanitization or validation of the text/content extracted from the PRD or prototypes before using it to drive code generation.
[EXTERNAL_DOWNLOADS]: The skill automatically installs external dependencies at runtime to facilitate visual verification.
Evidence: In scripts/render_page.py, the ensure_playwright function executes pip install playwright and playwright install chromium to ensure the environment is ready for screenshot capture.
Context: These are well-known, trusted tools maintained by Microsoft and the Chromium project respectively.
[COMMAND_EXECUTION]: The skill uses subprocesses to manage the development environment and executes user-generated scripts.
Evidence: scripts/render_page.py uses subprocess.run to install browser automation tools. Furthermore, Phase 7 of the skill explicitly instructs the agent to generate and then execute a start.sh script which may contain arbitrary commands for system setup and service management.
[DYNAMIC_EXECUTION]: The skill follows a pattern of generating executable content (shell scripts, backend code, frontend components) based on runtime data and then executing that content to verify the application.

glmv-prd-to-app