claude-code - 💡(How to fix) Fix [Question/Bug] Why did Claude 4.6 start using ①②③ⓐⓑⓒ for enumeration in Korean responses? [1 participants]

Summary

A behavioral change has been observed in Claude 4.6 (Opus and Sonnet) where Korean responses use Enclosed Alphanumerics (U+2460–U+24FF, e.g. ①②③, ⓐⓑⓒ) instead of standard ASCII markers (1. 2. 3. or -). This happens in lists and also in inline prose (e.g. ⓑ안 instead of B안).

한국어 원문: Claude 4.6 (Opus, Sonnet) 한국어 응답에서 열거형 표기를 1. 2. 3. 대신 ① ② ③, ⓐ ⓑ ⓒ 원문자(Enclosed Alphanumerics, U+2460~U+24FF)로 표기하는 행동이 관찰됨. 열거뿐 아니라 일반 문장 안에서도 B안 대신 ⓑ안 같은 표기가 나타남.

---

The Core Question

1. Was this intentional?
2. If yes — why, and how can users opt out?
3. If no — when did it start, and how will it be fixed?

한국어:

1. 이게 의도된 변경인가?
2. 의도된 거라면 이유가 뭐고, 어떻게 끌 수 있나?
3. 의도 안 된 거라면 언제부터 시작됐고 어떻게 고칠 건가?

---

Why It Matters

• These characters render as tofu (□) in many common terminal and editor fonts, degrading readability
• Copy-paste to HWP, older docs, or chat apps causes encoding issues
• The behavior was NOT present in earlier Claude versions — this is a usability regression for Korean users

한국어:

• 이 글자들은 많은 터미널/에디터 폰트에서 □(tofu)로 깨짐 → 가독성 저하
• HWP, 구형 문서, 채팅앱 등에 복붙 시 인코딩 문제 유발
• 이전 Claude 버전에는 없던 동작 → 한국어 사용자에게 사용성 회귀

---

Cross-Environment Evidence (Strongest Signal)

The identical symptom was observed across two completely different terminal/font setups, same machine, same user, same time window:

Environment A (Vera session):

• Ghostty: latest (just updated)
• Font: Sarasa Mono K (just switched from D2Coding minutes earlier)
• Result: ①②③ output observed

Environment B (Stock PM session, running concurrently):

• Terminal: unchanged (old config, untouched for weeks)
• Font: unchanged
• Result: ①②③ output observed — identical symptom

→ Rules out terminal as cause. → Rules out font as cause. → Rules out Ghostty version as cause. → Rules out OS as cause. → The only common factor between the two environments is the Claude model itself. → Strongly suggests a recent model-side output habit shift.

한국어 (교차 환경 증거 — 결정타): 같은 맥, 같은 사용자, 같은 시간대에 완전히 다른 두 터미널 환경에서 동일 증상 관찰됨.

• 환경 A (Vera): Ghostty 최신 업데이트 + Sarasa Mono K 방금 교체 → ①②③ 발생
• 환경 B (Stock PM, 동시 실행): 옛날 터미널 + 옛날 폰트 그대로 (몇 주간 안 건드림) → ①②③ 동일 발생 → 터미널/폰트/OS 모두 원인 아님. 유일한 공통 요소는 Claude 모델 자체.

---

No System Prompt Instruction

• No user instruction to use circled numbers
• No project CLAUDE.md mentioning these characters
• Behavior emerged without any explicit rule

한국어:

• 사용자 프롬프트에 원문자 사용 지시 없음
• 프로젝트 CLAUDE.md에 해당 문자 언급 없음
• 어떤 명시적 규칙도 없이 자발적으로 발현됨

---

Environment

• Claude Code: v2.1.101
• Models: claude-opus-4-6, claude-sonnet-4-6
• OS: macOS 25.3.0 (Darwin)
• Terminal: Ghostty (Environment A), unspecified older terminal (Environment B)
• Observed date: 2026-04-11
• Language: Korean responses

---

Attached Screenshots

Screenshots will be attached by the reporter after issue creation (7 images):

1. Natural Claude output using ⓑ inline — Claude wrote "ⓑ안이 부활" (meaning "Plan B is revived") in regular prose, not in a list context. Evidence of spontaneous model choice.
2. Meta-evidence (title + body with tofu) — screenshot of the reporter's own draft of this very bug report, rendered in their terminal, where the circled characters in the text appear as tofu boxes (□). The bug manifests inside the discussion of the bug.
3. Meta-evidence (English section with tofu) — same draft, English section.
4. Meta-evidence (Cross-environment paragraph with tofu) — same draft, the environment evidence paragraph, again with tofu rendering. 5-7. Rendering test matrix (blocks A-R) — a character test matrix the reporter generated to probe which Unicode blocks render correctly in their font. Enclosed Alphanumerics (block A), Roman Numerals (B), and Hangul enclosed jamo (Q) show as tofu; other blocks render fine. Demonstrates the font/rendering failure at the glyph level.

Together these form a complete chain of evidence:

• The model's choice to use the characters (screenshot 1)
• The rendering failure they cause (screenshots 5-7)
• The meta-observation of both happening simultaneously (screenshots 2-4)

---

Request

1. Confirm whether this is intentional behavior
2. If intentional: provide the reasoning and a documented opt-out mechanism (system prompt directive, config flag, or similar)
3. If unintentional: treat as a regression and fix in the next model update

한국어 요청:

1. 의도적 행동인지 확인 요청
2. 의도적이라면 이유 설명 + 끄는 방법 제공
3. 의도 안 된 거라면 회귀 버그로 처리 및 수정

---

Reporter note

This report was drafted in Korean by a Korean user and translated bilingually for international triage. The reporter's primary language is Korean. Please feel free to respond in either language.

이 리포트는 한국어 사용자가 작성했으며, 국제 트리아지를 위해 영어 병기했습니다. 한국어/영어 어느 쪽으로 답변하셔도 괜찮습니다.