AGI-in-MD: Using Markdown-structured Prompts to Enhance Reasoning Capabilities of Large Language Models

This article introduces the innovative prompt engineering method proposed by the AGI-in-MD project—mapping cognitive prompts into Markdown format, leveraging its hierarchical structure and semantic markers to enhance the reasoning capabilities of large language models, and improving their performance in tasks like code understanding, creative generation, and system analysis. This method provides a new approach for effective interaction with large language models.

Prompt engineering has evolved from early simple instructions to complex few-shot learning, chain-of-thought, and role-playing techniques. As model capabilities improve, the degree of structure in prompts has an increasingly significant impact on reasoning quality. Recently, the "agi-in-md" project on GitHub proposed mapping cognitive prompts into Markdown format to enhance model performance through clearer information organization.

Why Markdown is Suitable for Structured Prompts

Markdown has advantages such as clear hierarchical structure, rich semantic markers, wide compatibility, and good human readability, which are highly aligned with the needs of prompt engineering.

Structured Mapping of Cognitive Prompts

AGI-in-MD maps human cognitive processes (problem decomposition, element identification, relationship establishment, etc.) into Markdown structures: using heading levels to decompose problems, lists to enumerate elements, tables/nested lists to express relationships, and quote blocks/code blocks to mark reasoning processes.

Based on Markdown, structured prompts show advantages in multiple scenarios:

• Code Analysis and Review: Use headings to define review dimensions, lists to enumerate issues, and code blocks to present improvement suggestions;
• Creative Generation and Brainstorming: Hierarchical structure guides divergent-convergent thinking;
• System Analysis and Architecture Design: Heading levels express system component relationships, tables record interface dependencies, and Mermaid flowcharts show interactions;
• Knowledge Organization and Learning: Clear structure generates organized knowledge summaries.

Designing high-quality structured prompts requires following these guidelines:

1. Consistency Principle: Maintain a unified structural style;
2. Progressive Unfolding Principle: From high-level overview to details;
3. Visual Separation Principle: Use dividers/blank lines to distinguish phases;
4. Meta-information Annotation Principle: Explain task objectives and output formats at the beginning/end.

Markdown-structured prompts can be combined with other techniques:

• Chain of Thought: Embed "think step by step" prompts to guide reasoning;
• Role Playing: Use quote blocks to define role backgrounds + structured content to expand tasks;
• Few-shot Learning: Use Markdown code blocks/tables to provide examples and maintain consistent formatting.

When using Markdown-structured prompts, note the following limitations:

• Over-structuring Risk: Complex structures may confuse the model;
• Model Differences: Different models have varying degrees of understanding of Markdown;
• Context Length Limitation: Complex structures consume more tokens, so a balance between completeness and conciseness is needed.

AGI-in-MD represents a new thinking in prompt engineering—prompts are an art of information organization. Through Markdown-structured expression, more effective communication with models can be established, unlocking their reasoning potential. As LLM capabilities improve, the importance of prompt engineering increases; mastering structured prompt design is an essential skill for AI developers, and Markdown is an elegant starting point.