The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill accepts user input via the --prompt flag in several scripts (logo/generate.py, cip/generate.py, icon/generate.py). This input is interpolated directly into prompt templates sent to Gemini AI models. While this presents an attack surface where a user could attempt to influence the AI's generation process, the impact is limited to the generated image or SVG output.
[INGESTION POINTS]: User-provided prompts via CLI arguments.
[BOUNDARY MARKERS]: None detected in the script prompt templates.
[CAPABILITY INVENTORY]: File system writes (saving images/SVGs), network access (Google GenAI API).
[SANITIZATION]: Input strings are not sanitized before interpolation.
[DYNAMIC_EXECUTION]: The skill dynamically generates HTML files for presentations (scripts/cip/render-html.py) and social media photos (references/social-photos-design.md). These files are then intended to be rendered by a browser or captured as screenshots. This is a core functional requirement of the design workflow.
[DATA_EXPOSURE]: The Python scripts include a load_env function that reads .env files from the user's home directory (~/.claude/.env and ~/.claude/skills/.env) to retrieve the GEMINI_API_KEY. This is standard practice for secure credential management in this environment.

design