How to Reuse the Same Prompt Across ChatGPT, Claude, Gemini, and Ollama

Published March 23, 2026 · Updated March 24, 2026 · 3 min read

Developing a prompt for one provider and then needing it to work across several others is a common situation — for benchmarking models, for multi-provider architectures, or for switching providers as pricing or quality changes. The core instructions are portable; the JSON wrapper isn't.

What changes between providers

Three structural differences matter:

System prompt placement. OpenAI: inside messages array as {"role":"system"}. Anthropic: top-level "system" parameter. Gemini: "system_instruction" object. Ollama: inside messages array like OpenAI, but model support varies.

Role names for AI responses. OpenAI and Anthropic and Ollama: "assistant". Gemini: "model".

Content structure. OpenAI, Anthropic, Ollama: "content": "text". Gemini: "parts": [{"text": "text"}].

The conversation content — the actual instructions, examples, and constraints — requires no changes. Only the container format changes.

Writing prompts that port easily

Portable prompts separate the invariant parts (instructions, examples, constraints) from the structural parts (role names, content wrappers). Write the core of the prompt as clean text first, then wrap it in whichever format the target provider expects.

Avoid relying on provider-specific capabilities in the prompt logic. Tool calling syntax differs significantly between providers. Vision input structure differs. Some prompt patterns that work reliably on GPT-4o — like specific JSON output schemas — may require minor rewording to produce equivalent results on Claude or Gemini.

A practical conversion workflow

Develop the prompt in your primary provider
Use a prompt template converter to produce the target format
Test with 5–10 representative inputs, including edge cases
Adjust system instructions if output quality differs in ways that matter

The prompt template converter handles all three structural changes — system prompt placement, role renaming, and content wrapping — automatically. What it doesn't handle is model-specific behaviour differences that might require prompt wording adjustments.

Models have different defaults: the same prompt may produce different results across providers because models have different defaults and different instruction-following characteristics. Claude follows detailed system instructions very precisely. GPT-4o tends to be efficient on structured extraction. Gemini performs well on tasks with large amounts of context. Test with real inputs after converting.

When portability matters less

If you're building a single-provider production application, over-engineering for portability adds complexity for no immediate benefit. Portability matters when you're actively benchmarking multiple providers, building a multi-provider architecture, or in an early phase where the provider choice isn't final.