Structured context engineering: YAML and Markdown outperform JSON for LLM context

Source type	Practitioner
Confidence	Medium
Last verified	2026-03-20
Topics	claude-code

Insight: A systematic study of 9,649 experiments across 11 models and 4 formats (YAML, Markdown, JSON, TOON) found that frontier models (Opus 4.5, GPT-5.2, Gemini 2.5 Pro) benefit significantly from filesystem-based context retrieval, while open-source models show less convincing results. A "grep tax" was observed: Token-Oriented Object Notation (TOON), designed for minimal tokens, actually caused models to spend more tokens iterating due to format unfamiliarity.

Sources

• Simon Willison — "Structured Context Engineering for File-Native Agentic Systems" (2026-02-09)

Related: existing entry "Context engineering for AI-assisted development" in external/claude-code.md — CORROBORATES