The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to construct and execute shell commands that incorporate strings derived from user input and external analysis results. This pattern is vulnerable to command injection if the agent does not properly escape shell metacharacters.
Evidence: Shell command patterns in references/mode-banner.md, references/mode-batch.md, references/mode-product-main.md, and references/mode-reference-gen.md interpolate variables like [提示词], [活动名称], and [产品具体描述] directly into arguments for nano_banana2.py and render_template.py execution.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external URLs via browser screenshots, which are then analyzed and used to influence the agent's logic for prompt generation and script execution.
Ingestion points: The take_screenshot tool is used in references/mode-reference-gen.md (Step 1) to capture content from arbitrary user-provided URLs.
Boundary markers: No boundary markers or "ignore embedded instructions" delimiters are specified for the ingested content.
Capability inventory: The skill has the capability to execute shell commands via Python scripts in sibling directories (nano-banana2, ecommerce-designer) as detailed in the reference files.
Sanitization: No sanitization or validation of the analyzed design language is mentioned before it is used to generate prompts for the image generation scripts.

design-pipeline