The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of user-supplied text prompts and reference images, creating a standard surface for indirect injection.
Ingestion points: Raw user prompts and local image files provided via the --input flag in scripts/generate.ts.
Boundary markers: The references/prompt-guide.md instructs the agent to rewrite and narrate prompts, providing a manual normalization layer, though no programmatic delimiters are enforced.
Capability inventory: The skill invokes external AI APIs (Google, OpenAI, xAI), performs file system writes for image output, and executes sub-scripts via the bun runtime.
Sanitization: Relies on the agent's pre-generation rewriting and the safety filters implemented by the third-party image generation providers.
[DYNAMIC_EXECUTION]: The main generation script employs dynamic module loading to resolve the plugin's root directory across different installation environments.
Evidence: scripts/generate.ts uses import() with computed paths to find resolve-root.ts and utility modules by searching the local file system (e.g., ~/.claude/plugins/installed_plugins.json). This behavior is consistent with the vendor's architectural patterns for path resolution and does not involve untrusted data.

generate-image