The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes user-supplied prompts from files and standard input, which are then passed to external AI models. This represents a standard surface for indirect prompt injection common to AI-integrated tools.
Ingestion points: The skill accepts untrusted data through the --promptfiles flag, standard input, and via reference image paths provided to the --ref flag in scripts/main.ts.
Boundary markers: The skill does not implement specific boundary delimiters or safety instructions when interpolating these inputs into the API request payloads.
Capability inventory: The skill possesses the ability to read local files (readFile), write images to the disk (writeFile), and make network requests to AI providers (fetch).
Sanitization: No input sanitization or validation of the prompt content is performed before transmission to the AI provider.
[EXTERNAL_DOWNLOADS]: The skill connects to official AI provider endpoints to perform its intended function. These are well-known services and the interaction is necessary for the skill's operation.
Evidence: Network requests are made to dashscope.aliyuncs.com, generativelanguage.googleapis.com, and api.openai.com in scripts/providers/dashscope.ts, scripts/providers/google.ts, and scripts/providers/openai.ts.
[SAFE]: The skill follows established best practices for a CLI tool, such as using environment variables for secret management and performing file operations only on user-specified paths.

baoyu-image-gen