The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/_codex.py uses subprocess.run to execute an external Python script (codex_imagegen.py) expected to be part of the omnimedia skill. While the execution uses a list-based argument structure to prevent shell injection, it establishes a dependency on the presence and integrity of scripts in the user's home directory (~/.claude/skills/omnimedia/).
[DYNAMIC_EXECUTION]: The skill performs several dynamic generation and execution tasks:
scripts/poster/cluster.py uses scikit-learn and numpy for data clustering, then uses an LLM to synthesize design attributes which are written to local CSV files.
scripts/cip/render-html.py generates HTML presentations at runtime, embedding images as Base64 strings.
scripts/icon/generate.py relies on an LLM to produce raw SVG markup based on user descriptions.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data that is interpolated into LLM system prompts without sanitization, creating a vulnerability to indirect prompt injection.
Ingestion points: CLI arguments such as --brand, --prompt, and --logo (file path), as well as local images processed by scripts/poster/analyze.py via the --input-dir flag.
Boundary markers: Absent. System prompts in analyze.py, generate.py (logo/CIP/icon), and cluster.py do not utilize delimiters or specific instructions to ignore malicious commands embedded in user-supplied strings.
Capability inventory: The skill can write to the local file system (images, JSON, CSV, HTML), perform network operations via Google's GenAI APIs, and execute local commands through subprocess.run.
Sanitization: Absent. User input is directly formatted into design brief and generation templates without validation or filtering.
[EXTERNAL_DOWNLOADS]: The documentation in SKILL.md and reference files recommends the installation of external Python packages (google-genai, pillow) and a system tool (codex) via Homebrew to enable full functionality.

marketing-design