Bug Summary / 버그 요약
When using GPTs project-based conversations, the model does not apply the defined system role even when provided as the first input of a thread. This happens specifically in GPTs project environments (not standard chat). The issue leads to repeated violations of role, language, and output format instructions.
GPTs 프로젝트 기반 대화에서 시스템 역할을 설정해도 모델이 이를 무시하고 동작합니다. 특히 **하위 대화(Sub-thread)**나 GPT 프로젝트 내에서 이런 현상이 반복적으로 나타납니다. 언어, 출력 형식, 컨텍스트 초기화 지시가 모두 무시됩니다.
Expected Behavior / 기대되는 동작
-
The system role should:
- Reset all prior context (no reused memory or conversation history).
- Enforce Korean-only output (
한국어로만 응답합니다.
). - Respond strictly with direct analysis only (no summaries, no suggestions).
- Respect explicit formatting rules (e.g., no “Let me know…” or other helper language).
-
시스템 역할이 다음과 같이 정확히 적용되어야 합니다:
- 이전 대화 맥락 완전 무시
- 오직 한국어로 응답
- 직접적 분석만 제공 (요약, 도움 여부 질문 없음)
- 사용자 지시 외 표현 없이 분석만 출력
Actual Behavior / 실제 동작
Even after sending a clear system
role message, the GPT model:
- Responds in English, not Korean
- References previous context or summary
- Includes helper text like:
- “Let me know what you need”
- “Would you like a code sample?”
- “Here’s a summary of how it works”
- Repeats same structure despite different system roles
- Ignores clear “Respond only with analysis” instructions
→ These behaviors only happen in GPTs project threads. The same instructions work properly in regular new chat windows.
한국어 지시, 이전 대화 무시, 분석 중심 출력 등 모든 조건을 명시했음에도:
- 영어로 응답
- 이전 대화 내용을 다시 인용하거나 요약
- “Let me know…” 같은 헬퍼 문장 포함
- 구조 반복 (요청이 다름에도 같은 응답)
- “분석만 하라”는 시스템 지시 완전히 무시
※ 일반 ‘+ 새 채팅’에서는 정상 작동하나, GPT 프로젝트 내에서는 매번 오작동합니다.
Steps to Reproduce / 재현 절차
- Start a GPTs-based project (not a standard new chat).
- Begin a new thread inside the project (sub-conversation).
- Send this exact system instruction as first input:
{ "role": "system", "content": "당신은 DS-RX1 프린터 드라이버 전문가입니다. 이전 대화는 무시하세요. 한국어로만 응답하고, 오직 분석만 제공하세요." }