craft-human-agent-languagelisted

# Craft Human-Agent Language Pick the **language, structure, and output contract** for any text artifact that humans, agents, or both will read. Core method is model-agnostic: durable principles live here; dated measurements live in `references/` and must be re-measured after model releases. ## When to use - Write, translate, or review a prompt, README, ADR, update, or instruction that humans and/or agents consume. - Choose between zh / en / colloquial, or between prose / list / XML / code / JSON. - Switching model family or tokenizer and need a fresh language/format decision. ## Method (top-down, 6 layers) ### 0. Frame State purpose and audience before writing: - Purpose: `inform` | `decide` | `execute` - Audience: `human` | `agent` | `both` - Constraints: token cost, correctness, parseability ### 1. Structure before prose - Headings → bullets → numbered steps → code/JSON blocks. - Use XML tags to separate instructions from data (attention anchors + injection boundary). - One idea per line. No paragraph walls. - **Examples over descriptions**: one worked example beats one paragraph of explanation. ### 2. Plain language floor (ISO 24495-1) - Relevant: only what the reader needs; delete fluff. - Findable: headings, lists, numbering (ISO 2145 style). - Understandable: short active sentences; controlled vocabulary; one meaning per term. - Usable: callable steps; schema for agent consumers. - Open `references/iso-plain-language.md` for formal or long-lived artifa