jaymay549/ai-model-selector
Overview
This skill helps engineers and product teams pick the right LLMs, set API parameters, and implement calls across major providers. It summarizes model strengths, cost tradeoffs, context limits, common gotchas, and ready-to-use code templates. Use it to reduce cost, latency, and integration risk when building AI features.
How this skill works
It inspects your primary requirement (coding, reasoning, long context, throughput, embeddings, or multimodal) and maps that to recommended models and configurations. It highlights provider-specific parameters, pricing tiers, context windows, and parameter compatibility gotchas. It also provides concise code templates for OpenAI, Anthropic, Google, and DeepSeek to accelerate implementation.
When to use it
- Selecting an LLM for a new feature (chat, coding, reasoning, or multimodal)
- Configuring API parameters to avoid provider-specific errors
- Choosing cost-optimized pipelines for high-volume or batch workloads
- Designing RAG/embedding pipelines and choosing embedding models
- Deciding which model to use for very long documents or large context windows
Best practices
- Match model to task: use inference-focused models for throughput and flagship models for hardest reasoning or coding
- Avoid unsupported params for newer models (use reasoning.effort, text.verbosity, max_completion_tokens where required)
- Use cascading: route trivial requests to cheap models and hard problems to flagship models
- Enable batch API and prompt caching to cut costs on non-urgent workloads
- Validate long-context pricing tiers and test token accounting on target models
Example use cases
- Agentic coding assistant: Claude Sonnet 4.5 or GPT-5.2-Codex for code generation and reviews
- Math and scientific reasoning: GPT-5.2 (xhigh) or Claude Opus 4.5 for high-accuracy proofs and stepwise solutions
- Long-document ingestion: Gemini 2.5 Pro or Claude Sonnet 4.5 (1M beta) for multi-hundred-thousand token contexts
- High-volume chatbots: use Gemini Flash variants or Claude Haiku 4.5 for low latency and cost
- Embeddings and RAG: use Voyage family for highest-quality embeddings; use budget small embedding models for cost-sensitive pipelines
FAQ
Do not send temperature, top_p, or max_tokens. Use provider-native controls such as reasoning.effort, text.verbosity, and max_completion_tokens.
How do I optimize cost for large-scale non-urgent workloads?
Use batch APIs (many providers offer ~50% off), prompt caching to reduce charged inputs, and model cascading to push simple queries to cheaper models.
