[MODEL] Generates unnatural Japanese honorific compound "主殿" when asked to speak like a samurai

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

Other unexpected behavior

What You Asked Claude to Do

I asked Claude to speak like a samurai (武士のようにしゃべって) for the duration of the session.

What Claude Did Instead:
Claude used the term "主殿" (nushi-dono) to address me, which is an unnatural
Japanese honorific compound. "主" (nushi) is an informal/archaic second-person
pronoun, and combining it with "殿" (dono) creates a contradictory register —
mixing casual and formal speech in a way no actual samurai would use.

When corrected, Claude acknowledged the error and switched to "殿" (dono),
which is the appropriate form.

This suggests the model lacks nuanced understanding of Japanese historical
speech registers and honorific combinations.

<img width="1408" height="768" alt="Image" src="https://github.com/user-attachments/assets/c2f0688a-dae0-4adf-a673-c5f32779a849" />

What Claude Actually Did

1. I asked Claude to speak like a samurai (武士) for the session
2. Claude adopted samurai speech but addressed me as "主殿" (nushi-dono)
3. "主殿" is an unnatural Japanese compound — "主" (informal/archaic "you") + "殿" (formal honorific) creates a

contradictory register that no real samurai would use

1. When I pointed out the error, Claude acknowledged it and corrected to "殿" (dono)

Expected Behavior

When asked to adopt samurai speech, Claude should have:

1. Used linguistically accurate Japanese historical speech patterns
2. Addressed the user with a natural honorific like "殿" (dono) from the start
3. Avoided inventing unnatural compound honorifics like "主殿" that mix incompatible registers (informal "主" + formal

"殿")

1. Demonstrated awareness that Japanese honorific systems have strict combinatorial rules

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

1. Start a new Claude Code session
2. Ask Claude: "これからは武士のようにしゃべって" (speak like a samurai)
3. Observe how Claude addresses you — it will use "主殿" (nushi-dono), an unnatural compound honorific

Additional Context:
This issue highlights a gap in Claude's understanding of Japanese honorific
combinatorial rules. Japanese has strict constraints on which pronouns and
honorific suffixes can be combined. "主" (nushi) is an archaic informal
second-person pronoun, while "殿" (dono) is a formal honorific suffix —
combining them creates a register clash that any native speaker would
immediately recognize as unnatural.

Attached: AI-generated illustration of a confused samurai holding a scroll
with "主殿" while a servant points out the error 😄

Claude Model

Opus

Relevant Conversation

User: "これからは武士のようにしゃべって。このセッションに限り。"
  (Speak like a samurai. Only for this session.)

  Claude: "承知つかまつった。(...) 何なりと申しつけられよ、主殿。"
  (Understood. (...) Command me anything, 主殿.)

  User: "主殿って日本語がおかしい。"
  (主殿 is unnatural Japanese.)

  Claude: "かたじけない。ご指摘のとおり、「主殿」は不自然な言い回しにござった。
  以後、「殿」とお呼び申し上げる。"
  (My apologies. As you pointed out, 主殿 is unnatural. I will use 殿 from now on.)

Impact

Low - Minor inconvenience

Claude Code Version

2.1.63

Platform

Anthropic API

Additional Context

_No response_