The same prompt, three models: why structure beats wording

People swap prompts between AI tools the way they swap recipes, then wonder why the result is worse. The content usually transfers fine. The structure often does not.

Here is one plain request, and what it looks like written for three different models.

The original

write a blog post about remote work

Seven words, no audience, no length, no angle, no format. Every model will answer it, and every answer will be forgettable in the same way.

Written for ChatGPT

## Role
You are an experienced business writer.

## Task
Write a blog post about remote work for [audience].

## Requirements
1. 900-1100 words
2. One concrete example per section
3. No statistics unless I provide them

## Output format
Markdown, one H2 per section, ending with three takeaways.

Markdown headings act as boundaries. ChatGPT treats each headed section as its own instruction, and numbered requirements survive much better than the same requirements written as a sentence.

Written for Claude

<role>You are an experienced business writer.</role>

<context>
Audience: [audience]
What they already believe: [assumption]
</context>

<instructions>
Write a blog post about remote work.
</instructions>

<constraints>
- 900-1100 words
- One concrete example per section
- If a claim needs a statistic I have not given you, write [source needed]
</constraints>

Tags make boundaries unambiguous, which matters most when the prompt carries a lot of pasted material. With a long document inside a context tag, Claude is noticeably less likely to mistake your material for your instructions.

Written for Gemini

**Role:** Experienced business writer
**Task:** Blog post about remote work for [audience]
**Constraints:** 900-1100 words; one concrete example per section; no invented statistics
**Output:** Markdown, H2 per section, three takeaways at the end

Short labelled blocks, and the output shape stated explicitly. Gemini tends to sprawl when the shape is left open and tightens up considerably when it is not.

What actually changed

Notice what stayed the same across all three: the role, the audience, the length, the example requirement, and the rule about invented statistics. That is the content of a good prompt, and it is model-independent. What changed is only syntax.

This is why we build optimization in two steps. The model writes the parts of a prompt — role, objective, context, constraints, output format — and then an adapter renders those parts in the target's conventional shape. The parts are the thinking; the syntax is mechanical, and mechanical things should be deterministic.

Agents are a different job

Prompts for coding agents like Claude Code are not writing prompts at all. They work in a repository, so the useful structure is goal, context, requirements, approach and verification — how the work will be checked matters more than tone or word count. A framework designed for marketing copy is the wrong tool there, which is why we do not apply one.

The short version

• Content transfers between models. Structure does not.
• ChatGPT: Markdown headings, numbered requirements.
• Claude: XML-style tags, especially with long pasted material.
• Gemini: short labelled blocks, explicit output shape.
• Coding agents: goal, context, requirements, verification.