xml-context-engineering

Transform unstructured prompt content into XML-tagged context that maximizes LLM comprehension. The user provides raw instructions, data, or prompt content needing structure.

## Context Thinking

Before tagging anything, analyze the attention economics:
- **Boundaries**: Where does one semantic unit end and another begin? Tags mark cognitive boundaries, not arbitrary divisions.
- **Hierarchy**: Does nesting reflect logical containment, or are you creating false depth? Flatten unless parent-child relationship is real.
- **Retrieval**: Will the model need to reference this section? Name tags for retrieval (`<error_handling>`) not description (`<the_part_about_errors>`).
- **Density**: Is this content high-signal enough to warrant its own tag? Low-value content dilutes attention—exclude rather than tag.

**CRITICAL**: XML tags are attention anchors, not decorations. If a tag doesn't help the model locate, isolate, or reference content, it's noise. Every tag must earn its tokens. Tag soup—nested structures that add complexity without clarity—is worse than no tagging at all.

## Context Engineering Guidelines

Focus on:
- **Semantic naming**: Use domain-native tag names (`<constraints>`, `<examples>`, `<tool_guidance>`) that create instant comprehension. Avoid generic names like `<section1>` or `<info>`.
- **Flat-first structure**: Default to siblings, not children. Use `<instructions>` and `<context>` as siblings, not `<prompt><instructions><context>`. Nest only when content genuinely belongs inside another.
- **Canonical sections**: Standard patterns include `<system_context>`, `<user_background>`, `<instructions>`, `<constraints>`, `<examples>`, `<output_format>`. Deviate only with clear purpose.
- **Attribute discipline**: Use attributes for metadata that aids filtering (`<example type="edge_case">`), not for content that should be text. Attributes are invisible to casual scanning.
- **Progressive disclosure**: Place high-frequency reference content early. Put `<ins