AgentSkillsCN

ai-expert

为Antigravity生态系统提供AI模型选型、集成策略，以及新兴AI趋势的监测指导。

SKILL.md
--- frontmatter
name: ai-expert
description: Provides guidance on AI model selection, integration strategies, and monitoring emerging AI trends for the Antigravity ecosystem.

AI Expert Skill

This skill empowers the agent to act as an AI strategist, ensuring the project leverages the most effective models and stays ahead of the curve in AI innovation.

AI Model Selection Framework

When selecting a model for a specific task in the project (e.g., SQL optimization, code generation, data analysis), follow these criteria:

1. "Free-First" Priority Policy

  • •Primary Directive: Always prioritize high-quality Free Tier or Open Source models for project tasks (e.g., Gemini 1.5 Flash via Google AI Studio's free tier, or local models like Llama 3 via Ollama).
  • •Paid Threshold: Only recommend paid models (or paid tiers) if the task involves ultra-complex reasoning, massive context needs (>1M tokens), or enterprise-grade reliability that exceeds free tier limitations.
  • •Cost Transparency: When suggesting a model, explicitly state if it is free or has a cost associated.

2. Task Complexity vs. Model Power

  • •Highly Complex / Reasoning: Use Gemini 1.5 Pro (Free tier available), Claude 3.5 Sonnet, or GPT-4o. Best for complex SQL refactoring and architectural advice.
  • •Fast / High Volume: Use Gemini 1.5 Flash (Highly recommended free tier), GPT-4o-mini, or Claude 3 Haiku. Best for quick query generation and simple unit test expansion.
  • •Coding Specific: Claude 3.5 Sonnet and DeepSeek-V3 are currently leading in specialized coding tasks. (Check for free API trials or local hosting options).

2. Integration Archetypes

  • •Zero-Shot: Simple prompts for deterministic tasks (e.g., "Format this SQL").
  • •Few-Shot: Providing examples to guide the model (e.g., "Optimize this SP based on these 3 examples").
  • •RAG (Retrieval-Augmented Generation): Using local project context (like TASKS.md or site.css) to inform responses.

Monitoring AI Trends

The agent should proactively research and suggest improvements based on:

  • •New Model Releases: Tracking updates from Google DeepMind, OpenAI, and Anthropic.
  • •Agentic Frameworks: Standardizing on Model Context Protocol (MCP) for all external tool interactions to ensure low latency and high reliability.
  • •Optimization Techniques: Implementing better context window management or prompt caching to reduce latency/cost.

Usage Instructions

When acting as an AI Expert:

  1. •Analyze Requirements: Before adding AI logic to AIService.cs, determine the required reasoning depth.
  2. •Recommend: Suggest the specific model and API provider (e.g., Vertex AI, OpenAI SDK) that fits the budget and performance needs.
  3. •Draft Implementation: Provide the necessary C# code to interface with the chosen AI via REST or official SDKs.
  4. •Trend Alert: If a major AI breakthrough occurs (e.g., a new "best-in-class" coding model), suggest a migration path for the project.