[MODEL] Inconsistent output quality / Ignoring instructions (overfitting and inappropriate repetition of Korean vocabulary)

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 ignored my instructions or configuration

What You Asked Claude to Do

I asked Claude to respond in Korean using natural, context-appropriate vocabulary.
I also added explicit instructions to CLAUDE.md prohibiting certain inappropriate
words, and removed those words from all referenceable context and code.

What Claude Actually Did

1. Claude repeatedly used inappropriate or awkward vocabulary regardless of context.
2. Even after the problematic words were removed from all referenceable context and code,

Claude soon began using the same words frequently again.

1. Adding explicit prohibitions to CLAUDE.md had no effect — Claude kept using the

same words despite the documented instructions.

1. When this vocabulary usage was pointed out, the model excused itself by claiming

it was "behavior to save tokens."

Main inappropriate/awkward words observed:

• "박음" (bak-eum), "박힘" (bak-him): derivatives of "박다", which are slang. Used

consistently and repetitively instead of varied, context-appropriate expressions.

• "박제" (bak-je): could be replaced with various words that better fit the situation,

but only "박제" is used over and over.

• "정공" (jeong-gong): replaceable with diverse expressions such as "알맞은 (suitable),"

"정석 (standard/orthodox)," "권장되는 (recommended)," yet collapsed into one word.

Ambiguous vocabulary usage that leaves room for misinterpretation:

• "별" (byeol): indiscriminately unified into "별" in situations that require distinguishing

"별도 (separate)" from "개별 (individual)."

• "안" (an): indiscriminately unified into "안" in situations that require distinguishing

negation ("not ~") from spatial relation ("inside"), making the meaning ambiguous.

Expected Behavior

Claude should have:

1. Selected diverse and precise Korean vocabulary appropriate to the context.
2. Used standard, natural expressions instead of slang or awkward wording.
3. Disambiguated meanings that diverge — e.g., "별도/개별 (separate/individual)" and

"negation/spatial-inside" — rather than collapsing them into one ambiguous word.

1. Respected the prohibitions documented in CLAUDE.md, and not reused words that were

removed from context/code or explicitly prohibited.

Files Affected

N/A — this is not a file-modification issue, but a vocabulary-selection problem
that occurs during response text generation.
Note: the problematic words recur even after being removed from context and code,
and even after being explicitly prohibited in CLAUDE.md.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Use Claude Code in Korean for tasks such as code explanation, review, or documentation.
2. Observe the frequency of words like "박음", "박힘", "박제", "정공", "별", "안" in the responses.
3. Add explicit prohibitions for these words to CLAUDE.md, and remove them from context/code.
4. As the conversation continues, the same words reappear frequently despite the CLAUDE.md instructions.

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.153

Platform

Anthropic API

Additional Context

Rust projects
After a few conversations (5~20% of 1M context)