Why LLMs default to Markdown
Ask ChatGPT, Claude, or almost any other chat-based model for a structured answer, and it will format it in Markdown by default — headings, bold text, numbered steps, fenced code blocks. That's not an accident. Markdown signals structure with very few extra characters: a heading costs one # and a space, a list item costs one - and a space. The equivalent in HTML — <h2>, </h2>, <li>, </li> — spends far more tokens to say the same thing, and every extra token is money and latency at scale.
Markdown is also extremely well represented in the text these models were trained on. GitHub READMEs, Stack Overflow answers, technical documentation, and countless forum posts all use it, so models have seen an enormous number of examples of how headings, lists, and code fences map onto real structure. That familiarity makes Markdown a syntax models can both generate reliably and parse back out of their own (or a user's) text without ambiguity.
Markdown in prompts
The same structural clarity helps on the way in, not just the way out. A prompt broken into clearly labeled sections — say a ## Context heading, a ## Task heading, and a ## Constraints heading — gives the model an explicit map of what each part of the prompt is for, instead of leaving it to infer that from one undifferentiated paragraph.
- Use a bulleted or numbered list for instructions you want followed point-by-point — models tend to address list items individually rather than blending them together.
- Wrap any code, log output, or text the model should treat as literal (not rewrite or summarize) in a fenced code block. The triple-backtick fence is an unambiguous boundary most models respect.
- Use a table when you're giving the model several examples with the same shape — it reads as structured data rather than prose to summarize.
Converting documents for RAG
Teams building retrieval-augmented generation (RAG) pipelines routinely convert source documents — Word files, Confluence or Google Docs pages, scraped HTML — to Markdown before chunking and embedding them. Two things make that conversion worthwhile. First, Markdown keeps the structural signal (headings, lists, tables) that makes for good, coherent chunk boundaries — a heading is a natural place to start a new chunk. Second, it drops the visual noise: inline styles, class names, tracking attributes, and layout markup that inflate token counts without adding any meaning a retrieval system or a model can use.
If you're building a pipeline like this, HTML to Markdown and Word to Markdown handle exactly that first conversion step, entirely in your browser — useful when the source documents themselves shouldn't leave your machine before they're indexed.
What is llms.txt?
llms.txt is a proposed convention — modeled on the long-standing robots.txt — for a plain Markdown file published at a site's root that lists its most important pages and documents in a short, structured format. Where robots.txt tells crawlers what they may index, llms.txt is aimed at language models and AI agents doing retrieval at inference time: it's a curated map of a site's content, written in the format those models already parse best.
MarkdownLab publishes its own llms.txt for exactly this reason — a plain-text index of the tools and reference pages on this site, kept up to date as the tool set grows.
Practical patterns
- Use headings to establish hierarchy in both prompts and documents you want a model to reason over section by section.
- Prefer a real Markdown table over hand-aligned ASCII art for tabular data — it parses cleanly and takes fewer tokens.
- Always tag fenced code blocks with a language (
```python,```json) — it costs nothing extra and gives the model an explicit hint about the content's syntax. - Avoid deeply nested lists where possible; some parsers and prompt-processing pipelines flatten nesting past one or two levels.
- Use explicit link syntax (
[text](url)) rather than a bare URL when the link text itself carries meaning you want preserved.
Where to go next
New to the syntax itself? Start with What is Markdown or the full Markdown cheat sheet. Ready to write? Open the editor and try it live.