AI Translation Skill: From Simple Prompts to Agent Workflows

Translation is more complex than it seems. A simple prompt like 'translate this to Chinese' is easy, but a universal skill must handle varied inputs (single sentences, long documents, Markdown files), different target languages, and quality requirements. Key stages: 1) Pre-reasoning models: use two-step (literal then natural) or three-step (literal, review, natural) translation for quality. 2) With reasoning models: use 'rewrite' instead of 'translate' to give the model freedom for idioms and sentence restructuring. 3) Agent era: workflow with analysis, prompt generation, chunking by Markdown structure, parallel subagent translation, and overall proofreading. Save all intermediate results as files for iteration. Parallel translation requires consistency: pre-analyze terminology and strategy into a shared prompt file, so each subagent gets the same context. Split shared context (terminology, principles) into a file and task instructions (which file, where to save) as parameters. This makes prompts traceable and editable. Iterate by using the skill on real content, reviewing output, and feeding issues back to the agent for improvement.