Per-Model Prompting Guides

The Solution: Speak Each Model's Dialect

A per-model prompting guide is the practice of adapting the same task to the conventions each LLM family was trained on, instead of reusing one generic prompt everywhere. The underlying model architecture is similar across providers — a transformer predicting the next token — but the training data, instruction tuning, and recommended formatting differ. Those differences are why a prompt that shines on one model can feel mediocre on another. Think of it as speaking the same language in different dialects: the grammar is shared, but the idioms, punctuation, and politeness norms are not.

How it works

Each provider documents formatting that aligns with how the model was tuned. Anthropic recommends wrapping prompt sections in XML-style tags such as <document> and <instructions>, because Claude was trained on data where that structure carried meaning, and it follows the system prompt closely. OpenAI's GPT models lean on markdown headers, native function calling, and JSON mode for structured output. Google's Gemini is strongly multimodal and benefits from placing images near the start of the prompt, plus search grounding for fresh facts. Open-source models (Llama, Qwen, DeepSeek) expect an exact chat template with special tokens like <|im_start|> / <|im_end|>; skip the template and the model may ignore your system message entirely. The context window also varies widely — from a few thousand tokens on small open models to 200K–1M on frontier ones — so prompt length is part of the dialect too.

When to use it, tradeoffs, and a worked example

Adapt per model whenever output quality, cost, or reliability matters in production — and especially when you support several providers behind one feature. The main tradeoff is maintenance: keeping parallel prompt variants is more work than one shared template, and providers change defaults over time, so guides drift. A common pitfall is the "one prompt fits all" assumption, plus copying XML scaffolding onto GPT (where markdown reads more cleanly) or dropping a chat template on an open-source model. Worked example: say you extract an invoice total from a PDF. On Claude you would put the file inside <document> tags and ask for the answer inside <total> tags. On GPT you would request JSON mode and define a function returning {"total": number}. On Gemini you would attach the image first, then the instruction. Same task, three dialects — and measurably better results than forcing all three through one prompt.

Model-Specific Prompt Formats

• Claude (Anthropic): XML tags for structure, extended thinking, 200K context, prefilled responses, strong system prompt adherence
• GPT-4 / GPT-5 (OpenAI): JSON mode, function calling, markdown formatting preferred, instruction prioritization (later instructions win)
• Gemini (Google): True multimodal (images first in prompt), 1M+ token context, search grounding, structured prompt templates boost accuracy 40%
• Open-Source (Llama, DeepSeek, Qwen): Chat templates required (im_start/im_end), explicit formatting, shorter prompts work better, model-specific system prompt formats
• Common Pitfall: One Prompt Fits All: A prompt optimized for GPT-4 may underperform on Claude by 20-30% because Claude expects XML structure, not markdown headers. Always adapt to the model.

Try It Yourself!

Explore the interactive comparison below to see how the same task is prompted differently for each model and learn their unique features.