claude-code - 💡(How to fix) Fix Bug: LLM occasionally switches response language from Chinese (Traditional) to Korean or Japanese

Bug Description

When the system prompt / CLAUDE.md instructs Claude Code to always respond in 繁體中文台灣用語 (Traditional Chinese - Taiwan), the model occasionally switches to Korean or Japanese instead, without any user instruction to do so.

Steps to Reproduce

1. Set language instruction in CLAUDE.md or system prompt: ``` Always respond in 繁體中文台灣用語. Use 繁體中文台灣用語 for all explanations, comments, and communications with the user. ```
2. Have a long conversation session (especially one involving multi-step coding tasks or context compression).
3. Observe that at some point Claude Code starts responding in Korean or Japanese instead of Traditional Chinese.

Expected Behavior

Claude Code should consistently respond in Traditional Chinese (繁體中文台灣用語) throughout the entire session, regardless of session length or context compression events.

Actual Behavior

Responses spontaneously switch to Korean or Japanese mid-session. The switch appears to happen more frequently:

• After context window compression
• During tasks that involve multi-agent or multi-step workflows
• When processing code or tool results that contain CJK characters from mixed sources

Environment

• Tool: Claude Code CLI
• Language instruction: CLAUDE.md system-level instruction
• Affected languages observed: Korean (한국어), Japanese (日本語) instead of Traditional Chinese (繁體中文)

Additional Context

This is likely a model-level issue where the language instruction is not robustly maintained when the context contains mixed CJK (Chinese/Japanese/Korean) content, or when context compression occurs. The model may be inferring language from surrounding CJK characters rather than following the explicit instruction.

Workaround

None currently. User must manually remind the model to switch back to Chinese in each affected response.