The Agent Skills Directory

[PROMPT_INJECTION]: The scripts/build_design_prompt.py script interpolates data from an external JSON file (provided via --idea-json) directly into a prompt template without sanitization.
Ingestion points: The title, description, and category fields from the user-provided idea JSON are read in main() and passed to build_prompt().
Boundary markers: The prompt uses Markdown headers (e.g., - **Description**: {description}) but lacks explicit delimiters or instructions to the LLM to ignore potentially malicious commands within that data.
Capability inventory: The generated prompt is intended for use with the Claude CLI (claude -p), which in the provided workflow (SKILL.md Step 3) is granted powerful tools: Read, Edit, Write, Glob, and Grep. An attacker providing a malicious description could potentially cause the agent to write or modify files outside the intended scope.
Sanitization: There is no escaping or validation of the input fields before they are embedded in the final prompt string.

skill-designer