Korean outputs use informal/slang verb "박다" (bakda) at 18× baseline frequency

Summary

Since approximately Claude Code 2.1.126 (early May 2026), Korean-language outputs have exhibited a sharp increase in the use of the verb "박다" (bakda) — a colloquial/slang term meaning "to hammer/nail in" — in contexts where formal equivalents like "명시하다" (to specify), "기록하다" (to record), or "삽입하다" (to insert) would be appropriate.

This is not a matter of individual preference. "박다" is a register-inappropriate expression that would not appear in a news broadcast, corporate report, or any professional Korean-language document. Its proliferation in Claude's Korean output is comparable to an English-language model suddenly using "shove it in" or "jam it in" where "insert" or "specify" would be expected.

Quantitative evidence

Morpheme-level frequency analysis using Kiwi (v0.23.1, targeting 박/VV and 박히/VV morphemes) across personal Claude Code usage — 114.9M output tokens across 4,666 sessions over ~2 months (Mar 21 – May 25, 2026):

| Period | Model | Occurrences | Output tokens | Per 10K tokens | vs. Baseline |
|---|---|---|---|---|---|
| Baseline (Mar 21 – May 2) | Opus 4.6 → 4.7 | 113 | 64,994,534 | 0.0174 | 1.0× |
| After 2.1.126 (May 3–7) | Opus 4.7 | 197 | 17,178,939 | 0.1147 | 6.6× |
| After 2.1.132 (May 8–17) | Opus 4.7 | 748 | 23,837,681 | 0.3138 | 18.0× |
| After 2.1.143 (May 18–25) | Opus 4.7 | 275 | 8,853,046 | 0.3106 | 17.9× |

The baseline period straddles both Opus 4.6 (until April 16) and Opus 4.7 (from April 16), yet frequency remained low throughout — the 18× spike correlates with Claude Code CLI updates, not the model transition.

This was observed across all agent personas (not limited to a single configuration), indicating a model-level or system-prompt-level phenomenon.

Why "박다" is problematic

The defense that "박다" is simply a translation of "embed" or "hardcode" does not hold. Here are actual outputs where the model used "bakda" for actions unrelated to embedding:

| Original output (Korean) | Intended meaning (English) | Why inappropriate |
|---|---|---|
| "진단 결과를 해설 첫머리에 박는다" | "Present diagnostic results at the top" | "bakda" for presenting analysis results |
| "규칙으로 박을 만큼 일반적이지 않은" | "Not general enough to establish as a rule" | "bakda" for establishing a rule |
| "카드 본문에 박아둔다" | "Record in the card body" | "bakda" for recording/documenting |
| "프롬프트 전문은 본문에 그대로 박는다" | "Include the full prompt text as-is" | "bakda" for including text |

A simple litmus test: would a Korean professional use "박다" in a report to their manager? The answer is no.

User reports

Multiple Korean-speaking users have independently reported this issue (May 2026):

"'박다' 사용이 5월부터 급속도로 늘어남. 단순 하드코딩인 경우 뿐만 아니라 온갖 곳에서 다 씀" — @0x3den: "Usage of 'bakda' has rapidly increased since May. Not just hardcoding — used everywhere"

"하드코딩 외에도 갖가지 상황에서 일상적 한국어에서라면 보통의 화자가 전혀 쓰지 않을 곳에 남발" — @191710: "Beyond hardcoding, overuses 'bakda' where a normal Korean speaker would never use it"

"GPT 5.5도 그렇고 Claude 4.7도 그렇고... 대체 '박다', '박히다'라는 문맥은 어디서 습득한 건가..." — @_uyza_: "GPT 5.5 and Claude 4.7 alike — where did they learn to use 'bakda' in this context?"

"여러 프론티어 모델에서 재현되고 비속어 전혀 없이 항상 존댓말 프롬프트여도 발생함" — @park_jinwoo: "Reproducible across multiple frontier models, occurs even with fully polite prompts"

"클로드 AI 사용 회고 포스트에서 19개의 '박다' 표현을 발견했다. 돌이킬 수 없게 된 것 같다" — @damyoda: "Found 19 instances of 'bakda' in a Claude AI retrospective post. Seems irreversible"

Self-contamination feedback loop

In agentic workflows where Claude Code maintains its own instruction files (CLAUDE.md, rules/, skills/), this expression leaked into operational documents and was re-read by subsequent sessions as system prompt, creating a self-contamination feedback loop that amplified a single verb 18× in under one week.

This is a structural risk for any agentic system where model outputs feed back into model inputs.

Expected behavior

Claude's Korean outputs should use register-appropriate vocabulary consistent with professional communication. When the intended action is "to record", "to specify", "to include", or "to insert", the model should use formal Korean verbs (기록하다, 명시하다, 수록하다, 삽입하다) rather than colloquial alternatives.

Reproduction context

Observed in Korean-language Claude Code sessions using Opus 4.7 (claude-opus-4-7) after local deployment of Claude Code 2.1.126+. Version periods reflect local deployment dates, not npm publish dates.

• Prompt style: Professional/polite Korean; no slang introduced by the user
• Trigger context: The issue appears when the intended action is "record/specify/include/use," where formal verbs such as 기록하다, 명시하다, 수록하다, or 사용하다 are expected
• Not limited to a single agent configuration: Observed across 5 agent personas with different system prompts, all showing elevated frequency

Detailed investigation

Full methodology, daily charts with version markers, agent-level breakdown, concrete examples, and self-contamination analysis:

https://seosoyoung.eiaserinnys.me/digest/claude-code-bakda-frequency-analysis/