Markdown in Prompts

Why Structure Helps the Model

A wall of text requires the model to segment your prompt mentally before it can act on it. When you use markdown headers to divide your prompt into labelled sections — Task, Context, Constraints, Input — the model can parse each section independently and maintain clearer mental boundaries between what you want, what you're providing, and what limits you're imposing. Structure also makes it harder for elements to bleed into each other: your constraints won't accidentally be interpreted as part of the task description.

Headers for Section Separation

Use ## headers to create clear sections in multi-part prompts. Common sections include: ## Task (what you want the model to do), ## Context (background the model needs), ## Constraints (limits on the output), ## Input or ## Article or ## Data (the raw material to work with), and ## Output Format (how to structure the response). This structure makes long prompts dramatically easier to read and iterate on.

Code Fences and Tables

Use triple-backtick code fences with language identifiers to include code samples, JSON, or structured data in your prompt — this signals that the enclosed text should be treated as literal content rather than instruction: \\\python ... \\\``. If you're providing tabular data, markdown tables can help the model parse relationships between rows and columns more accurately than a CSV dump.

When to Avoid Markdown in Prompts

Some contexts render markdown as raw text rather than formatting — plain text APIs, terminal output, certain mobile apps. In these contexts, markdown in prompts (and in outputs) will appear as literal asterisks and hashes. Similarly, if you're building a voice interface, markdown structure in the output is unhelpful. Check how your interface renders text before relying heavily on markdown formatting in either prompts or expected outputs.