[MODEL] Assistant occasionally outputs Simplified Chinese characters mid-conversation despite fully Traditional Chinese (Taiwan) context

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

Ongoing conversation conducted entirely in Traditional Chinese (Taiwan usage). I explicitly told Claude multiple times, in this same conversation, not to reply in Simplified Chinese.

What Claude Actually Did

Claude wrote a reply containing the sentence "內容跟我读到的一致" — using the Simplified character 读 instead of the correct Traditional character 讀 — despite the entire conversation being in Traditional Chinese and having been explicitly asked repeatedly not to do this. This was not a one-off: it recurred periodically across the conversation.

Expected Behavior

Claude should consistently use only Traditional Chinese characters throughout the entire conversation, with zero Simplified Chinese characters, especially after being explicitly asked not to use them.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

「內容跟我读到的一致」

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.274 (Claude Code)

Platform

Anthropic API

Additional Context

This is a recurring issue for Traditional Chinese (Taiwan/Hong Kong) users, not a one-time fluke — it has happened repeatedly across long conversations, even after being explicitly corrected within the same conversation. For Traditional-Chinese-speaking users this significantly undermines trust, since it happens despite clear, repeated instructions.