Markdown Symbols - The key to speaking "LLM-ese"

6th May 2025 — 2 minute read

When to Use Markdown in Prompts permalink

Use it when you want to:

Structure complex information clearly
Create reusable templates or step-by-step flows
Trigger “thinking mode” (e.g., writing in outline, code, or comparison format)

Why This Works: Training Exposure permalink

LLMs were trained on massive corpora that include:

GitHub repos
StackOverflow
Docs like READMEs and API guides
Markdown-based websites like Notion, Reddit, blogs, etc.

So when you use common Markdown formatting, you’re speaking the model’s native visual language.

High-Impact Markdown Symbols (Strongly Learned Patterns) permalink

Symbol	Markdown Use	Why It Matters to LLMs
`#`, `##`, `###`	Headings	Triggers structured thinking; segments instructions or sections clearly.
`*`, `-`, `+`	Bullets	Recognized as lists — helps the model organize thoughts and outputs.
`1.`, `2.`	Numbered lists	Supports step-by-step reasoning or procedural thinking.
`bold` or `__bold__`	Emphasis	Suggests importance or labels — model often echoes this in output.
`_italic_` or `italic`	Subtle emphasis	Often used to signal asides, tone, or nuance.
`> blockquote`	Quoting	Interpreted as citing another voice or referencing prior text.
```	Code blocks	Very strong cue: model switches to structured syntax, preserves formatting, uses technical language.
`---`	Horizontal rule	Seen as a topic or section break — triggers a shift in focus.

Moderate-Impact Symbols permalink

Symbol	Use	LLM Behavior
`[]()`	Links	Recognized in context (e.g., tutorials), but rarely “clicked” — treated as source metadata.
`![alt](image)`	Image embeds	Model recognizes this is an image placeholder — it might respond descriptively.
`\` (escape)	Escape character	Rarely changes LLM behavior unless in code contexts.
HTML inside markdown	Tables, anchors, divs	Sometimes parsed correctly, especially in documentation, but inconsistent in free text.

Low-Impact / Cosmetic Markdown permalink

Symbol	Use	LLM Behavior
` ` or raw HTML escapes	Fine-tuning spacing	Typically ignored or stripped in LLM response.
Line breaks (`\n`)	Formatting	Helps readability, but model doesn’t always preserve layout unless told to.