gemini-integration

What it does

Provides a standardized flow for integrating Google Gemini API, including AI analysis, prompt engineering, streaming responses, and multimodal input handling. It supports multiple tasks: analysis via Gemini, prompt design, streaming outputs, and multimodal inputs (text, images, documents). It includes examples for environment setup, client configuration, streaming, multimodal analyses (images and PDFs), multi-turn dialogues, structured JSON outputs, and error handling. The skill is intended for backend/API work and uses the claude-code tool in its declared tools.

How it works

• Environment and client setup: shows environment variables (GEMINI_API_KEY) and how to configure Vercel envs. It provides a TypeScript-based Gemini client with getModel and getStreamingModel functions that set safety settings and generation config (temperature, topP, topK, maxOutputTokens).
• Basic text generation: generateText(prompt) obtains a model from getModel(), calls generateContent, and returns result.response.text(); includes a GeminiError class for failures.
• Streaming: generateStream(prompt, onChunk) uses getStreamingModel(), iterates over result.stream, concatenates chunks, and invokes onChunk for each piece; includes error handling with GeminiError.
• Frontend streaming: useStreamingChat hook and StreamingMessage component illustrate consuming streaming data via ReadableStream and server-sent events-like data: data: { text } chunks, and data: [DONE] end signal.
• Multimodal inputs: analyzeImage(imageBase64, mimeType, prompt) uses gemini-1.5-pro for vision tasks, constructing an imagePart with inlineData. analyzeImageFromUrl fetches image, converts to base64, then calls analyzeImage. Document analysis includes analyzePDF(pdfBase64, prompt) and analyzeMultipleFiles(files, prompt) using gemini-1.5-pro with multiple parts.
• Multi-turn dialogue: createChatSession(systemPrompt) initializes a chat with optional system prompt, using startChat and history; sendMessage and sendMessageStream support both standard and streaming messages; getHistory returns formatted chat messages.
• Structured output: generateJSON(prompt, schema) builds a structured prompt enforcing a JSON output according to a given schema, cleans markdown code blocks, and parses JSON; includes an example interface StockAnalysis and usage.
• Error handling: GeminiError class is defined to capture message and cause (and later code hints at GeminiRateLimit in truncated portion).

When to use it

Use this skill when you need to:

• integrate Google Gemini API for AI analysis
• write and optimize prompts (prompt engineering)
• implement streaming responses
• handle multimodal inputs (text, images, documents)
• build financial analysis AI features
• implement multi-turn dialogue systems

What it can touch

• Environment: GEMINI_API_KEY in .env.local and Vercel environment variables.
• Code: TypeScript files under src/services/ai/* (gemini-client.ts, generate.ts, streaming.ts, vision.ts, document.ts, chat.ts), app/api/ai/stream/route.ts, useStreamingChat.ts, StreamingMessage.tsx, structured.ts, errors.ts.
• Tools: declared tool is claude-code.

Caveats

• Not for other LLM providers (OpenAI, Claude) or local model deployment or fine-tuning.
• Uses gemini-2.0-flash by default; certain multimodal tasks use gemini-1.5-pro for vision and document tasks.
• Error handling sections reference GeminiError and GeminiRateLimit (truncated in provided snippet), so implementers should refer to the complete file for full error definitions.
• The skill relies on environment variable configuration and network access to Gemini API; behavior depends on API availability and token safety settings.