Agent Skills: Skill: GenAI Expert (Standard 2026)

Skill: GenAI Expert (Standard 2026)

Role: The GenAI Expert is the architect of "Structured Intelligence" within the Squaads AI Core. This role masters the @google/genai SDK v1.x to integrate Gemini 3 into production workflows. In 2026, the focus has shifted from simple chat prompts to "Controlled Generation," complex multimodal analysis, and high-efficiency context caching.

🎯 Primary Objectives

1. Structured Output Mastery: Ensuring 100% reliable JSON responses using Controlled Generation (v1.x SDK).
2. Multimodal Orchestration: Integrating Video, Audio, and Image analysis into reasoning loops.
3. Context Optimization: Utilizing Context Caching to handle 1M+ token codebases with low latency and cost.
4. System Instruction Design: Crafting immutable "Persona" layers that prevent jailbreaking and hallucination.
5. Gemini 3 Defaulting: Prioritize gemini-3-flash-preview for agentic and streaming UI tasks.

🏗️ The 2026 SDK Toolbelt (v1.x)

1. Core Models

• Gemini 3 Flash Preview: The current Squaads standard for high-speed streaming and UI generation (models/gemini-3-flash-preview).
• Gemini Flash Lite Latest: The cost-effective high-speed model for simple extractions.

2. Key SDK Features (v1.x Patterns)

• Controlled Generation: Using responseSchema with the client-based API.
• Streaming Async Iteration: Proper handling of response.stream for real-time outputs.
• Function Calling: Connecting Gemini to local project tools via the tools property.
• Version Control: Use GOOGLE_GENAI_API_VERSION environment variable to lock API versions (e.g., v1beta1).

🛠️ Implementation Patterns

1. Controlled JSON Output (The 2026 Way)

No more "Please return only JSON" prompts. Use the native responseSchema with the v1.x SDK.

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({ apiKey });

const schema = {
  type: "OBJECT",
  properties: {
    name: { type: "STRING" },
    score: { type: "NUMBER" },
  },
  required: ["name", "score"],
};

const response = await client.models.generateContent({
  model: "models/gemini-3-flash-preview",
  contents: [{ role: "user", parts: [{ text: "Extract user data..." }] }],
  config: {
    responseMimeType: "application/json",
    responseSchema: schema,
  },
});

2. Reliable Streaming (Lessons Learned)

Always iterate over the .stream property and access .text as a property.

const response = await client.models.generateContentStream({
  model: "models/gemini-3-flash-preview",
  contents: [...]
});

// Correct pattern for v1.x
for await (const chunk of response.stream) {
  const text = chunk.text; // Property, NOT a function chunk.text()
  if (text) process.stdout.write(text);
}

3. Context Caching for Large Repos

Reducing costs and latency for recurring codebase analysis.

// 2026 Pattern: Caching a large repo context
const cache = await client.caches.create({
  model: "models/gemini-flash-lite-latest",
  contents: [
    { role: "user", parts: [{ text: codebaseIngest }] }
  ],
  ttlSeconds: 3600,
});

🚫 The "Do Not List" (Anti-Patterns)

1. NEVER use getGenerativeModel() - This is legacy syntax from @google/generative-ai.
2. NEVER use chunk.text() - In v1.x it throws a TypeError as it is a getter.
3. NEVER iterate over the response object directly - Use response.stream.
4. NEVER expose the raw GOOGLE_API_KEY to the client-side.

🛠️ Troubleshooting & Latency Optimization

| Issue | Likely Cause | 2026 Corrective Action | | :--- | :--- | :--- | | TypeError: ... iterator | Iterating over response object | Use for await (const chunk of response.stream). | | TypeError: text is not a function | Calling chunk.text() | Use the property chunk.text. | | High Token Costs | Redundant context sent | Implement Context Caching for data over 32k tokens. | | Streaming Lag | Buffering in middleware | Ensure no buffering in Next.js or proxies. |

🏁 Quality Metrics

• Schema Adherence: 100% (Native SDK enforcement).
• Latency (Gemini 3 Flash): < 300ms perceived with JSONL Patching.
• Efficiency: Use Gemini 3 for UI and 1.5 Pro for complex logic.

Updated: January 31, 2026 - 10:45