AI Chat Studio

Part of Agent Skills™ by googleadsagent.ai™

Description

AI Chat Studio provides a multi-LLM chat orchestration framework with 300+ assistant presets, intelligent model routing, and conversation management. The agent configures and manages interactions across multiple language model providers—OpenAI, Anthropic, Google, open-source models—selecting the optimal model for each task based on capability, cost, and latency requirements.

Not every task needs the most powerful model. A code review benefits from a reasoning-heavy model; a translation task runs well on a mid-tier model; a simple reformatting task wastes money on anything beyond a fast, cheap model. This skill implements intelligent routing that matches task characteristics to model capabilities, reducing cost by 40-60% while maintaining quality where it matters.

The 300+ assistant presets encode domain-specific system prompts, temperature settings, and output format constraints for common tasks: code generation, technical writing, data analysis, creative ideation, customer support, legal review, and more. Each preset is tested against a quality benchmark and tagged with the models it performs best on.

Use When

• Configuring multi-provider LLM access in an application
• Routing tasks to the optimal model by cost-quality trade-off
• Managing conversation history and context windows
• Deploying domain-specific AI assistants with curated presets
• Building chat interfaces with streaming responses
• Comparing model outputs for the same prompt across providers

How It Works

graph TD
    A[User Message] --> B[Task Classifier]
    B --> C{Task Type}
    C -->|Complex Reasoning| D[Claude 4 / GPT-4o]
    C -->|Code Generation| E[Claude 4 / Codestral]
    C -->|Translation| F[GPT-4o-mini / Gemini Flash]
    C -->|Simple Format| G[Haiku / Flash]
    D --> H[Apply Preset: System Prompt + Params]
    E --> H
    F --> H
    G --> H
    H --> I[Manage Context Window]
    I --> J[Stream Response]
    J --> K[Log Usage + Cost]

The task classifier analyzes the incoming message to determine complexity and domain, then routes to the most cost-effective model capable of handling it. Presets provide domain-specific system prompts and parameter tuning.

Implementation

interface ModelConfig {
  provider: "openai" | "anthropic" | "google" | "ollama";
  model: string;
  maxTokens: number;
  costPer1kInput: number;
  costPer1kOutput: number;
  capabilities: string[];
}

const MODEL_REGISTRY: ModelConfig[] = [
  { provider: "anthropic", model: "claude-sonnet-4-20250514", maxTokens: 8192,
    costPer1kInput: 0.003, costPer1kOutput: 0.015, capabilities: ["reasoning", "code", "analysis"] },
  { provider: "openai", model: "gpt-4o-mini", maxTokens: 4096,
    costPer1kInput: 0.00015, costPer1kOutput: 0.0006, capabilities: ["general", "translation", "format"] },
  { provider: "google", model: "gemini-2.0-flash", maxTokens: 8192,
    costPer1kInput: 0.0001, costPer1kOutput: 0.0004, capabilities: ["general", "fast", "multimodal"] },
];

interface AssistantPreset {
  id: string;
  name: string;
  systemPrompt: string;
  temperature: number;
  preferredModels: string[];
  tags: string[];
}

class ChatRouter {
  constructor(private models: ModelConfig[], private presets: Map<string, AssistantPreset>) {}

  route(message: string, presetId?: string): { model: ModelConfig; preset?: AssistantPreset } {
    const preset = presetId ? this.presets.get(presetId) : undefined;
    const taskType = this.classifyTask(message);

    const candidates = this.models.filter(m =>
      m.capabilities.some(c => taskType.requiredCapabilities.includes(c))
    );

    const selected = candidates.sort((a, b) => a.costPer1kInput - b.costPer1kInput)[0];
    return { model: selected, preset };
  }

  private classifyTask(message: string): { type: string; requiredCapabilities: string[] } {
    const lower = message.toLowerCase();
    if (lower.includes("debug") || lower.includes("refactor") || lower.includes("architect"))
      return { type: "complex", requiredCapabilities: ["reasoning", "code"] };
    if (lower.includes("translate") || lower.includes("rewrite"))
      return { type: "simple", requiredCapabilities: ["general", "translation"] };
    return { type: "general", requiredCapabilities: ["general"] };
  }
}

class ConversationManager {
  private history: Array<{ role: string; content: string }> = [];
  private maxContextTokens: number;

  constructor(maxContextTokens: number = 100_000) {
    this.maxContextTokens = maxContextTokens;
  }

  addMessage(role: string, content: string): void {
    this.history.push({ role, content });
    this.trimToContextWindow();
  }

  getHistory(): Array<{ role: string; content: string }> {
    return [...this.history];
  }

  private trimToContextWindow(): void {
    while (this.estimateTokens() > this.maxContextTokens && this.history.length > 2) {
      this.history.splice(1, 1);
    }
  }

  private estimateTokens(): number {
    return this.history.reduce((sum, m) => sum + Math.ceil(m.content.length / 4), 0);
  }
}

Best Practices

• Route simple tasks to cheaper models—80% of queries do not need frontier models
• Implement streaming responses for all chat interactions to improve perceived latency
• Trim conversation history from the middle, preserving the system prompt and recent messages
• Log model selection decisions alongside cost to optimize routing rules over time
• Test presets against a benchmark dataset before deploying to production
• Provide fallback models for every route in case the primary provider is unavailable

Platform Compatibility

Platform	Support	Notes
Cursor	Full	Multi-model configuration
VS Code	Full	Extension-based LLM access
Windsurf	Full	Built-in model routing
Claude Code	Full	Multi-provider support
Cline	Full	Model selection config
aider	Full	Multiple model backends

Related Skills

• Assistant Presets
• Workflow Orchestration
• Multi-Model Routing
• Knowledge Base Injection

Keywords

ai-chat multi-llm model-routing assistant-presets conversation-management streaming cost-optimization chat-studio

---

© 2026 googleadsagent.ai™ | Agent Skills™ | MIT License