The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted user input (such as titles and themes) and interpolates them directly into prompt templates for image generation. This presents an indirect prompt injection surface where a user could potentially attempt to influence the agent's behavior through the provided text. ● Ingestion points: User-supplied titles and themes processed via SKILL.md and references/presets.md. ● Boundary markers: The skill uses double quotes and specific delimiters like '' and '「<中文标题>」' to separate user text from prompt instructions, but does not explicitly instruct the agent to ignore directives within that text. ● Capability inventory: The skill is limited to text generation; it has no file system, network, or code execution capabilities. ● Sanitization: No explicit input validation or escaping mechanisms are defined for user-provided strings.
[NO_CODE]: The skill contains no executable scripts (Python, JavaScript, shell) or compiled binaries. It operates entirely through markdown instructions and static templates, significantly reducing its attack surface.

bauhaus-visual