AI Output Formats

The format in which an AI agent delivers its work. Format choice significantly affects comprehension, usability, and information density.

Common Formats

• Plain text — universal, fast; no structure, no interactivity
• Markdown — lightweight structure, readable raw; loses visual/spatial info, no interactivity
• HTML — rich, interactive, self-contained; slight overhead to generate
• JSON/structured — machine-readable, precise; poor human readability
• PDF — print-ready, fixed layout; not interactive, not editable

The HTML Case

From html-effectiveness: HTML is the pragmatic sweet spot for AI agent deliverables. A single self-contained file with minimal JS can convey:

• Side-by-side comparisons
• Collapsible sections for progressive disclosure
• Annotated diffs
• Interactive diagrams and flowcharts
• Keyboard-navigable presentations

Format Shapes Content

The same information in different formats is not equivalent. Spatial relationships, hierarchies, and comparisons that are natural in HTML are lossy when forced into linear text.

Implications for Agentic Workflows

• Agents should be explicitly instructed to use HTML when output is complex or visual
• Default to markdown only for simple, linear text responses
• Self-contained HTML files (no external deps) are ideal: portable, archivable