By model
How prompting differs across the major models
The same prompt can behave differently on different models. Here is the honest, general picture — with the reminder that specifics change fast, so always confirm against the official docs.
OpenAI (GPT family)
Strengths: Strong general reasoning and tool use; a router-style setup where asking it to “think hard” can trigger deeper reasoning.
Prompting: Try zero-shot first — it infers intent well from minimal context. Skip generic “think step by step” on the reasoning tiers. Use native structured output for JSON.
Anthropic (Claude family)
Strengths: Careful instruction-following on long contexts and strong tone control; supports explicit extended thinking.
Prompting: Place long documents near the top and the question at the end; use clear section structure (e.g. XML-style tags). Let extended thinking handle reasoning rather than hand-written chains.
Google (Gemini family)
Strengths: Very large context windows and capable reasoning tiers.
Prompting: Google’s guidance favors including a few examples over zero-shot, and placing your specific question after the data context. With huge contexts, placement matters more, not less.
Meta (LLaMA family)
Strengths: Open-weight models you can self-host and tune; behavior depends heavily on the variant and system prompt.
Prompting: Lean on a clear system prompt; structure matters because defaults vary by deployment. Test your prompt on your specific variant rather than assuming.