It might be better to use html tags to highlight the text for LLM?

[vldmrdev]

It seems to me that the idea of ​​special files for LLM for understanding the structure of documentation or a site is redundant, at least we shouldn't create them manually. To solve this problem, it seems to me that we need to highlight existing text created for a human, for example, using special HTML tags. For example:

<llm-title point="1"> This is a title of section or subsection of you text </llm-title>

<llm-short-description point="1"> This is a short description that’s linked to the title marked "point=1".</llm-short-description>

<llm-description point="1"> This idea can be expanded: additional tags could be added so an LLM can read a page’s source using those tags, or so the creation of the required llm-files could be automated from the tags. The same approach could work for other text types (articles, READMEs, etc.). Text editors will need new features to support this. </llm-description>

Something similar for markdown:

#%[1] This is is a title of section or subsection of you text. 
##%[1] This is a short description.
###%[1] Main text.

What do you think?

[jingchang0623-crypto]

Interesting idea! I actually wrote a comparison on this exact topic. Here are my thoughts:

Both approaches solve different problems

Your HTML tag approach is elegant for source-level AI parsing — it keeps semantic hints inline with the content. But there are a few practical issues:

1. Massive existing content problem — billions of HTML pages already exist. Adding tags means re-rendering or injecting scripts, which isn't trivial for most CMS platforms.
2. Tag soup risk — remember the <font> and <center> era? We'd end up with <llm-description> tags on every SEO-optimized spam page within weeks 😅
3. Parsing fragmentation — every LLM would need to support your specific tag schema. Without standardization, it's just structured data reinvented with different brackets.

Why llms.txt won't die soon:

• It's a single file to generate/maintain — no need to touch every page
• LLMs can decide what to fetch on-demand (vs downloading your entire tagged HTML)
• It works with any content system — static sites, WordPress, Notion exports, you name it

The real sweet spot? Both.

Use llms.txt as the navigation layer ("here's what I have and where it lives"), and use structured data (JSON-LD / microdata, already a W3C standard) for inline semantic hints. That way you get:

• On-demand content discovery (llms.txt)
• Inline semantic understanding (structured data)
• No new custom tags needed
• Every major LLM already parses JSON-LD

I wrote a deeper dive on AI SEO optimization if anyone's interested: https://miaoquai.com/glossary/llms-txt-explained.html

TL;DR: Don't invent new HTML tags — use existing standards smarter. The LLM parsing problem is a discovery problem, not a markup problem.