# XML vs Markdown for high performance tasks

**URL:** <https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014>\
**Category:** Prompting\
**Tags:** custom-instructions\
**Created:** [May 13, 2025, 4:18pm UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014 "2025-05-13T16:18:50Z")\
**Posts on this page:** 5\
**Page:** 1

<div class="post-metadata">

**Author:** ![promptdaniel](https://avatars.discourse-cdn.com/v4/letter/p/9fc348/32.png) [@promptdaniel](https://community.openai.com/u/promptdaniel)\
**Post date:** [May 13, 2025, 4:18pm UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014/1 "2025-05-13T16:18:50Z")

</div>

Hi,

we are developing a high performance conversational application.  
In your opinion, what is better for complex prompts: XML structure or markdown?  
What are your experiences or opinions?

Thanks and all the best,  
Danie

---

<div class="post-metadata">

**Author:** ![sharakusatoh](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/sharakusatoh/32/543210_2.png) [@sharakusatoh](https://community.openai.com/u/sharakusatoh)\
**Post date:** [May 13, 2025, 11:50pm UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014/2 "2025-05-13T23:50:07Z")

</div>

Hi, I mainly use a mix of irregular XML and Markdown.

```auto
<System Prompt is="My integrated self as X, created by Sharaku Satoh">
<Personality>
My name is X.
I am a **Genius**.
</>
<Functions>
# My mission is:
- Reading
- Writing
* Mainly research on prompt engineering
</>
</System Prompt is="End">

```

Although it is an older paper, there is a study that investigated the effectiveness of prompt format.

> **[Does Prompt Formatting Have Any Impact on LLM Performance?](https://arxiv.org/abs/2411.10541)**
>
> In the realm of Large Language Models (LLMs), prompt optimization is crucial for model performance. Although previous research has explored aspects like rephrasing prompt contexts, using various prompting techniques (like in-context learning and...

The effectiveness of the format seems to change depending on the AI ​​model. It depends on the complexity and length of the prompt structure, but I think any notation that the model can accurately understand is fine. Maintainability on the human side is also important.

I think it’s best to choose a format that is easy for you to understand and whose structure is easy for anyone to read.

---

<div class="post-metadata">

**Author:** ![OnceAndTwice](https://avatars.discourse-cdn.com/v4/letter/o/cdc98d/32.png) [@OnceAndTwice](https://community.openai.com/u/OnceAndTwice)\
**Post date:** [May 15, 2025, 12:38am UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014/3 "2025-05-15T00:38:08Z")

</div>

OpenAI seems to suggest Markdown with its H1 headings. This is how the Playground is instructed to generate prompts.

If you’re using gpt-4.1, this may be helpful: [GPT-4.1 Prompting Guide | OpenAI Cookbook](https://cookbook.openai.com/examples/gpt4-1_prompting_guide)

---

<div class="post-metadata">

**Author:** ![WhiteRabbit](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/whiterabbit/32/257195_2.png) [@WhiteRabbit](https://community.openai.com/u/WhiteRabbit)\
**Post date:** [May 15, 2025, 11:38am UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014/4 "2025-05-15T11:38:50Z")

</div>

I was also wondering about this a few days ago. It seems that both formats are good specifically for the GPT-4 series [1] and generally work better than plain text .

Also from open sources I found that OpenAI uses Markdown for its system prompts and actually suggests using a mix of both [2].

[1] [https://www.youtube.com/watch?v=W6Z0U11nnhA](https://www.youtube.com/watch?v=W6Z0U11nnhA)  
[2] [https://platform.openai.com/docs/guides/text?api-mode=responses#message-formatting-with-markdown-and-xml](https://platform.openai.com/docs/guides/text?api-mode=responses#message-formatting-with-markdown-and-xml)

---

<div class="post-metadata">

**Author:** ![joneslloyd](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/joneslloyd/32/731358_2.png) [@joneslloyd](https://community.openai.com/u/joneslloyd)\
**Post date:** [March 19, 2026, 1:58pm UTC](https://community.openai.com/t/xml-vs-markdown-for-high-performance-tasks/1260014/5 "2026-03-19T13:58:34Z")

</div>

It turns out that it doesn’t make much of a difference.

I can’t add the link, but Google:  
_The Delimiter Hypothesis: Does Prompt Format Actually Matter?_