DeepSeek Harness ships its agent preset with an English system prompt. The model therefore thinks and plans in English, then translates the final answer back into Chinese. For Chinese-speaking users this is not just aesthetics: the whole reasoning chain runs in a second language, and both quality and token budget pay for it.
The demand
- Official Discussion #320 (27 upvotes, 2026-08-13, the second-hottest Chinese-language thread): “Don’t just copy claudecode — we want our own Chinese agent.”
- The supply side, in numbers: of 11,838 indexed repositories, 68 touch i18n at all, and the most-starred of those has 106★. Nobody has shipped the most direct form — a preset whose system prompt is Chinese. Meanwhile 46% of the indexed repos carry Chinese descriptions. Enormous base, near-zero supply.
The gap
Existing work targets UI translation or model swaps. Nobody has touched the preset itself: take the official cordis/standard preset and make persona and instructions natively Chinese, so thinking, planning and self-checks run in Chinese end to end.
Proposed shape
An agent preset package (the .agent-presets/<id>/cordis.yml form). A preset directory is a first-class distribution unit in DSH; the user switches presets in settings and touches no config file.
Implementation notes
- Read the official preset’s full structure first (the host-composition vs session-preset split); do not just translate the system-prompt strings. Tool descriptions, approval copy and plan templates must switch language together, or the model bounces between Chinese and English mid-thought.
- Translation is not transliteration: constraint phrasing (“NEVER do X”) loses force when moved word-by-word into Chinese. Rewrite in Chinese instruction style, then regression-test in real sessions.
- Ship a “hybrid” option: reasoning in Chinese, code identifiers and error messages left in English — the compromise many users actually want.
The MVP
One preset directory plus a README: fork the standard preset, localize the persona, document the install. A day for v1, then iterate on wording from real sessions.
Risks
Prompt quality is subjective — publish your evaluation method (which tasks, what counts as “reasoning improved”) or the preset becomes another “feels nice” artifact. That is also the opening: whoever publishes the first public benchmark for Chinese presets owns the conversation.
FAQ
Doesn’t switching to a Chinese model solve this?
No. The model sets the capability ceiling; the system prompt sets the behavioral chain. Under an English prompt, even a Chinese model reasons in English first and translates back.
Does this conflict with UI-localization plugins?
No — different layers, and they compose.