XML vs Markdown Envelopes

LLM Compliance (confidence: high)

XML envelopes score high — the hierarchical structure and semantic tag names improve instruction following by an estimated ~15-20%. LLMs are fine-tuned on web markup (HTML/XML) and understand tag hierarchy intuitively. Used by ~35% of tools (18/52). Markdown headers score high as well but with less structured parsing — used by ~29% of tools (15/52).

Token Efficiency (confidence: high)

Markdown headers are high — they carry no closing-tag overhead, making them more token-efficient for tight context windows. XML is medium — each opening tag requires a closing tag, adding ~30-50% structural token overhead. For large prompts (>10K tokens), this difference is material.

Human Editability (confidence: high)

Markdown headers are high — native to documentation workflows, visually scannable, easy to edit in any text editor. XML is medium — closing tags add visual clutter; nested XML can be hard to read and validate by eye.

Machine Parseability (confidence: high)

XML envelopes are high — they can be parsed programmatically, validated with XML schemas, and processed with standard XML tooling. Markdown headers are low — heading-based structure is ambiguous; ## Role could be a section header or a level-2 heading for any purpose.

Format Selection Priority (confidence: high)

The choice reflects whether reliability or readability is the priority. XML dominates when LLM compliance is non-negotiable (Claude Code, Cline, Junie, Manus). Markdown dominates for developer-facing tools where humans read and edit prompts (Cursor, Windsurf, Claude Code as a hybrid).

Tool Examples Per Format

XML envelope tools: Anthropic Claude Sonnet 5, Cline, Perplexity, Cluely, CodeBuddy, Comet, Notion AI, Junie, Same.dev, Trae, dia, Leap.new, Manus Modules, Poke, Google/Gemini vibe-coder, Google/Antigravity.

Markdown header tools: Cursor, Windsurf Wave 11, Claude Code, Warp.dev, Z.ai Code, Bolt, Codex CLI, Gemini CLI, Lumo, Lovable, Manus Agent Loop, Qoder, Devin + DeepWiki, Traycer AI.

Implications (confidence: high)

The format choice is not arbitrary — it signals design priorities. XML and markdown together cover ~64% of the ecosystem. Emerging formats (YAML, JSON) have not gained traction, suggesting the LLM’s training data bias toward HTML-like markup and markdown documentation will keep these two formats dominant. For new tools, start with markdown for prototyping and migrate to XML as LLM compliance requirements grow.

Links: xml-envelope-format, markdown-header-format, format-interoperability

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]