MAGI vs llms.txt - when to use which?

[jenalysematejek42]

I'm relatively new to making content AI-friendly and I'm trying to understand the landscape. The README mentions llms.txt as a comparison point.

From what I understand:

• llms.txt is a single file at the root that tells LLMs about a site
• MAGI is a per-document format with rich metadata

Are they complementary or competing approaches? Could you use llms.txt to point to MAGI documents? Would love to hear from people who've evaluated both.

[kendanbano72]

From a strategy perspective, llms.txt solves discovery while MAGI solves processing depth. If you're building an AI-first knowledge base, you want both.

The footnote relationships in MAGI are something llms.txt can't replicate - cross-document linking with typed relationships is essential for building knowledge graphs, not just flat document retrieval.

[jenalysematejek42]

That makes a lot of sense. So the stack would be:

llms.txt          -> Site map for LLMs
*.mda (MAGI)      -> Rich per-document format  
ai-script blocks  -> Processing instructions
footnotes         -> Document relationships

Thanks for clarifying! I think a comparison guide in the docs would help other newcomers.

[h-leisure846]

I've been evaluating both for our documentation platform and here's my take:

llms.txt is great for site-level configuration — telling LLMs "here's what this site is about and how to navigate it." It's like robots.txt but for AI.

MAGI operates at the document level — each file carries its own context, processing instructions, and relationship graph. It's more like Schema.org markup but for markdown.

They're actually complementary rather than competing:

• llms.txt → site-level AI guidance
• .mda files → document-level AI context

We're planning to use both: llms.txt for our marketing site, MAGI for our technical documentation.

[louiswolf2]

Good framing. We use llms.txt for site-level discovery and .mda for the actual content; they layer well rather than competing.

[santoyaworobey2]

From a dev-rel perspective: llms.txt for the public surface, MDA for the sources of truth in the repo. That mental model has worked for us.