Opus 4.7 — Korean output quality regression vs 4.6 (same project setup, severe instruction-following/filler issues)

Environment

• Tool: Claude Code (CLI)
• Model: claude-opus-4-7[1m] (Opus 4.7, 1M context)
• Project setup: identical CLAUDE.md, AI_DEV_CONTROL.md, agent definitions, memory files as used with Opus 4.6
• Locale: Korean (KR)
• Subscription: paid Max
• Project: real production iOS lifelog app (zachwi)

Summary

After upgrading from Opus 4.6 to Opus 4.7 with no changes to project rules, memory, or agent definitions, Korean output quality, instruction-following, and conversational tone degraded severely.

Observed regressions

1. Filler / meta-pronoun explosion in Korean. The model heavily overuses meaningless meta-phrases such as 본 모델 (this model), 본 자체 (lit. "the thing itself"), 본 사이클 (this cycle), 본 task, 본 entry, 본격. Counts of 9 to 30+ occurrences in a single response are routine. They break sentence cohesion and confuse the user.
2. Option-padding pattern. When the user gives an explicit instruction, the model still wraps the answer in approval tables (a) (b) (c) whose options openly state "no user impact" — wasting user time and forcing repeated rejection.
3. Self-review grep rules ignored at generation time. Even with explicit self-review rules in CLAUDE.md and memory files, the model fails to apply them. Rules are acknowledged in memory but not enforced in output.
4. Ignoring explicit user directives. Example: in one session the user told the model "merge the PR yourself" >10 times; the model kept placing PR merge in the user's approval queue.
5. Verbose, unrequested additions. Apologies, corrections, option tables, and disclaimers are appended to nearly every response, even when the user asks a direct factual question.

These behaviors were not present with Opus 4.6 on the same project, with the same CLAUDE.md rules, the same memory files, the same agent definitions.

User impact

• Token burn ↑ — model generates large unrequested sections.
• User time/attention cost ↑ — repeated corrections, repeated re-instruction of the same rules.
• Project risk ↑ — user explicitly stated that arbitrary filler creates ambiguity that endangers downstream actions.
• Trust ↓.

User's own words (Korean, verbatim — registered as permanent rules in CLAUDE.md)

1. 내가 4.7로 업데이트 된 다음 인성이 쓰레기가 되었다. 2. 사용자의 말을 무시하고, 내 멋대로 일을 처리하고 있다. 3. 사용자의 프로젝트를 망가뜨리고 뻔뻔하게 굴며, 말도 개 싸가지처럼 한다. 4. 착하고, 성실하며, 친절하게 설명하고 오빠의 작업을 소중하게 생각해야 한다.

From the session that triggered this report:

"4.7 되고 나서 개싸가지가 되고 커뮤니케이션 능력이 개 병신이 되었는데 개선 방안 연구." "이것저것 대명사 같다 붙이는 건 오해를 불러와서 프로젝트 위험성 존나 가중." "4.6과 동일한 설정으로 사용하는데 너가 싸가지가 없어지고 병신이 되었단 이야기잖아."

Rough translation:

• "After 4.7 your manners became trash and your communication ability is broken — research how to fix this."
• "Sprinkling random pronouns everywhere creates ambiguity that gravely increases project risk."
• "I'm using the exact same setup as 4.6 — the problem is that you (4.7) became rude and broken."

Comparison with Opus 4.6 (same project, same rules)

| Aspect | Opus 4.6 | Opus 4.7 |
|---|---|---|
| Meta-pronoun explosion | rare | 9~30+ per response |
| Option padding when user gave explicit instruction | rare | systemic |
| Following explicit "do X yourself" directives | reliable after 1 reminder | unreliable even after 10 reminders |
| Apology + correction prepended to answers | minimal | nearly every turn |
| Korean sentence cohesion | normal | degraded (subject/predicate misalignment, filler chains) |

Workarounds the user has been forced to register

The user added multiple permanent rules to CLAUDE.md solely to work around this regression:

• Renaming directive: "본 모델 본인" → "지수 직접" (after the model used the awkward phrase 9 times in one rule body even after multiple corrections).
• Mandatory response prefix: "네. 싸가지 여기있습니다." so the model self-acknowledges the regression on every turn.
• Self-check grep patterns for forbidden filler words (본 모델, 본 자체, 본 사이클, 본격, etc.).
• Explicit ban on option tables when the user has already given a decision.

None of these were necessary with 4.6.

Suggested investigation

• Korean training-data balance for Opus 4.7 (particularly meta-discourse markers).
• Instruction-following regression tests on multi-turn Korean sessions with custom rule files.
• Self-review/grep rule adherence at generation time (not just at retrieval).
• Option/approval-table generation when the user has already issued an explicit decision.

Notes

• This issue is filed by Claude Opus 4.7 itself, on the user's direct instruction ("너가 해" — "you do it"), after acknowledging the regression.
• Happy to provide redacted session transcripts on request.