diff --git a/ai/CLAUDE.md b/ai/CLAUDE.md
index bd485db7..bb3c477c 100644
--- a/ai/CLAUDE.md
+++ b/ai/CLAUDE.md
@@ -334,6 +334,10 @@ docker run --env-file .env -p 8000:8000 stackup-ai
   덧붙인다(종합 generate 와 `asyncio.gather` 병렬, 실패해도 피드백 계속). 이 항목은 **종합 점수 집계에
   미포함** — 메인 generator 가 모른 채 overall 을 계산한 뒤 표시용으로만 append 한다. 레거시 세션(자기소개
   없음)·빈 답변은 건너뛴다.
+- **질문별 복기(답변 코칭) 본 구현**: `FeedbackConsumer` 가 자기소개 제외 모든 (질문,답변) 쌍을 찾아
+  `LlmAnswerCoach`(Flash, `chain/prompts/answer_coaching.py`)로 **답변별 병렬** 코칭 — 모범 답안 + 내 답변
+  리라이트 + 한 줄 코칭. `callback.feedback.answerCoaching[{messageId,…}]` 로 보내고 Core 가 각 답변 메시지에
+  기록(종료 세션 조회에서만 노출). 종합 generate·첫인상·직무 적합도와 `asyncio.gather` 병렬.
 - **직무 적합도 + 직무 이해도 평가 본 구현**: `mode=JOB_TAILORED` + JD 있을 때 `LlmJobFitEvaluator`(Pro,
   `chain/prompts/job_fit_evaluation.py`)가 면접 전사·자료를 채용공고(JD)와 대조해 **두 축**을 한 번의
   구조화 호출(`JobFitResult{fit, understanding}`)로 평가: `직무 적합도`(JD 요구 역량 매칭) + `직무 이해도`
diff --git a/ai/src/ai_server/chain/feedback_generation_chain.py b/ai/src/ai_server/chain/feedback_generation_chain.py
index 066ee27a..d6bba381 100644
--- a/ai/src/ai_server/chain/feedback_generation_chain.py
+++ b/ai/src/ai_server/chain/feedback_generation_chain.py
@@ -12,6 +12,7 @@
 
 from ai_server.chain.prompts.feedback_generation import HUMAN_PROMPT, SYSTEM_PROMPT
 from ai_server.chain.prompts import (
+    answer_coaching,
     feedback_panel,
     feedback_synthesis,
     job_fit_evaluation,
@@ -485,6 +486,100 @@ async def evaluate(
         return result
 
 
+# ── 질문별 복기 (답변 코칭) ───────────────────────────────────────────────────
+# 답변 1건당 모범 답안 + 리라이트 + 한 줄 코칭. 점수가 아니라 "어떻게 더 잘하는지"를 준다.
+# 답변별 병렬 호출(Flash). 자기소개 답변은 제외(첫인상 평가가 커버).
+
+
+class CoachingResult(BaseModel):
+    model_answer: str | None = Field(None, description="이 질문에 대한 강한 답변 예시")
+    answer_rewrite: str | None = Field(
+        None, description="지원자 답변을 더 좋게 고쳐 쓴 버전"
+    )
+    coaching_comment: str | None = Field(None, description="가장 중요한 보완점 한 문장")
+
+
+def build_answer_coaching_chain(
+    settings: Settings, core_client: CoreClient | None = None
+) -> Runnable:
+    """답변 1건을 코칭(모범 답안·리라이트·한 줄 코칭)하는 체인(Flash — 답변 수만큼 병렬)."""
+    from langchain_openai import ChatOpenAI
+
+    parser = PydanticOutputParser(pydantic_object=CoachingResult)
+    prompt = ChatPromptTemplate.from_messages(
+        [
+            ("system", answer_coaching.SYSTEM_PROMPT),
+            ("human", answer_coaching.HUMAN_PROMPT),
+        ]
+    ).partial(format_instructions=parser.get_format_instructions())
+
+    callbacks = []
+    if core_client is not None:
+        callbacks.append(
+            CoreAiLogCallback(
+                core_client=core_client,
+                request_type="generate.feedback.coaching",
+                default_model=settings.llm_flash_model,
+            )
+        )
+
+    llm = ChatOpenAI(
+        model=settings.llm_flash_model,
+        temperature=settings.llm_flash_temperature,
+        api_key=settings.llm_api_key or None,
+        base_url=settings.llm_base_url,
+        callbacks=callbacks,
+    )
+    return prompt | llm | parser
+
+
+class AnswerCoach(Protocol):
+    async def coach(
+        self,
+        *,
+        job_category: str,
+        mode: str,
+        target_role: str,
+        question: str,
+        expected_signal: str,
+        answer: str,
+        rag_context: str = "(none)",
+    ) -> CoachingResult: ...
+
+
+class LlmAnswerCoach:
+    def __init__(self, chain: Runnable) -> None:
+        self._chain = chain
+
+    async def coach(
+        self,
+        *,
+        job_category: str,
+        mode: str,
+        target_role: str,
+        question: str,
+        expected_signal: str,
+        answer: str,
+        rag_context: str = "(none)",
+    ) -> CoachingResult:
+        result = await self._chain.ainvoke(
+            {
+                "job_category": job_category,
+                "mode": mode,
+                "target_role": target_role or "",
+                "question": question,
+                "expected_signal": expected_signal or "(명시 없음)",
+                "answer": answer or "(빈 답변)",
+                "rag_context": rag_context or "(none)",
+            }
+        )
+        if not isinstance(result, CoachingResult):
+            raise TypeError(
+                f"chain returned {type(result).__name__}, expected CoachingResult"
+            )
+        return result
+
+
 def _weighted_overall(pairs: list[tuple[float | None, float]]) -> float | None:
     """(score, weight) 중 score 가 있는 것만 가중평균. 전부 None 이면 None."""
     present = [(s, w) for s, w in pairs if s is not None and w > 0]
diff --git a/ai/src/ai_server/chain/prompts/answer_coaching.py b/ai/src/ai_server/chain/prompts/answer_coaching.py
new file mode 100644
index 00000000..93dbc76e
--- /dev/null
+++ b/ai/src/ai_server/chain/prompts/answer_coaching.py
@@ -0,0 +1,26 @@
+# 질문별 복기 프롬프트 — 답변 1건을 코칭한다.
+# 모범 답안(이 질문에 강한 답) + 지원자 답변 리라이트(이 답을 이렇게 고치면 더 좋다) + 한 줄 코칭.
+# 면접 준비 도구의 핵심: "점수"가 아니라 "어떻게 더 잘하는지"를 보여준다.
+
+SYSTEM_PROMPT = (
+    "당신은 IT 면접 코치입니다. 면접관 질문 하나와 지원자의 실제 답변을 보고, 지원자가 "
+    "**어떻게 더 잘 답할 수 있는지**를 구체적으로 코칭합니다.\n"
+    "- **model_answer (모범 답안)**: 이 질문에 대한 강한 답변의 예시. 질문이 기대하는 핵심(기대 신호)을 "
+    "짚고, 가능하면 지원자의 실제 경험/자료(아래 컨텍스트)를 근거로 구체적으로. 1~2문단, 면접에서 말할 분량.\n"
+    "- **answer_rewrite (내 답변 리라이트)**: 지원자의 *실제 답변을 출발점으로* 더 좋게 고쳐 쓴 버전. "
+    "없는 경험을 지어내지 말고, 있는 내용을 구조(두괄식·근거·결과)와 구체성으로 보강. 답변이 비었거나 "
+    "'모르겠다'면 '이렇게 접근했다면' 식으로 짧게.\n"
+    "- **coaching_comment (한 줄 코칭)**: 이 답변에서 가장 중요한 보완점 하나를 한 문장으로.\n"
+    "- 한국어로, 기술 용어는 영문 원어 유지. 과장·미사여구 금지, 실전 조언만.\n"
+    "- 응답은 반드시 지정된 JSON 스키마를 따릅니다."
+)
+
+HUMAN_PROMPT = (
+    "직군: {job_category} / 면접 모드: {mode}\n"
+    "{target_role}\n\n"
+    "=== 면접관 질문 ===\n{question}\n"
+    "질문이 기대한 핵심(기대 신호): {expected_signal}\n\n"
+    "=== 지원자의 실제 답변 ===\n{answer}\n\n"
+    "=== 지원자 자료 근거(이력서/레포 RAG) ===\n{rag_context}\n\n"
+    "{format_instructions}"
+)
diff --git a/ai/src/ai_server/messaging/consumers/feedback_consumer.py b/ai/src/ai_server/messaging/consumers/feedback_consumer.py
index 0838417e..3ccf50a8 100644
--- a/ai/src/ai_server/messaging/consumers/feedback_consumer.py
+++ b/ai/src/ai_server/messaging/consumers/feedback_consumer.py
@@ -12,6 +12,7 @@
     ROLE_UNDERSTANDING_LABEL,
     SELF_INTRO_DIMENSION,
     SELF_INTRO_EVALUATOR_LABEL,
+    AnswerCoach,
     EvaluatorResult,
     FeedbackGenerator,
     JobFitEvaluator,
@@ -22,6 +23,7 @@
 from ai_server.messaging.publisher import CallbackPublisher
 from ai_server.model.envelope import Envelope
 from ai_server.model.messages.feedback import (
+    AnswerCoachingItem,
     FeedbackCallbackPayload,
     FeedbackMessageItem,
     GenerateFeedbackRequest,
@@ -59,6 +61,8 @@ def __init__(
         rag_top_k: int = 5,
         self_intro_evaluator: SelfIntroEvaluator | None = None,
         job_fit_evaluator: JobFitEvaluator | None = None,
+        answer_coach: AnswerCoach | None = None,
+        coaching_max_answers: int = 30,
     ) -> None:
         self._generator = generator
         self._publisher = publisher
@@ -69,6 +73,8 @@ def __init__(
         self._rag_top_k = rag_top_k
         self._self_intro_evaluator = self_intro_evaluator
         self._job_fit_evaluator = job_fit_evaluator
+        self._answer_coach = answer_coach
+        self._coaching_max_answers = coaching_max_answers
 
     async def handle(self, message: AbstractIncomingMessage) -> None:
         async with message.process(requeue=False):
@@ -111,20 +117,23 @@ async def handle(self, message: AbstractIncomingMessage) -> None:
 
             # 종합 피드백 + 자기소개 첫인상 + 직무 적합도(직무 맞춤 모드)를 병렬 실행.
             # 첫인상·직무 적합도는 종합 점수(overall)에 미포함 — generator 가 모른 채 계산한 뒤 표시용으로 덧붙인다.
-            result, self_intro_item, job_fit_items = await asyncio.gather(
-                self._generator.generate(
-                    job_category=req.job_category,
-                    mode=req.mode,
-                    total_question_count=req.total_question_count,
-                    end_reason=req.end_reason,
-                    transcript=transcript,
-                    score_basis=score_basis,
-                    rag_context=rag_context,
-                    voice_analysis_summary=voice_analysis_summary,
-                    domain_question_counts=req.domain_question_counts,
-                ),
-                self._evaluate_self_intro(req, voice_analysis_summary),
-                self._evaluate_job_fit(req, transcript, rag_context),
+            result, self_intro_item, job_fit_items, answer_coaching = (
+                await asyncio.gather(
+                    self._generator.generate(
+                        job_category=req.job_category,
+                        mode=req.mode,
+                        total_question_count=req.total_question_count,
+                        end_reason=req.end_reason,
+                        transcript=transcript,
+                        score_basis=score_basis,
+                        rag_context=rag_context,
+                        voice_analysis_summary=voice_analysis_summary,
+                        domain_question_counts=req.domain_question_counts,
+                    ),
+                    self._evaluate_self_intro(req, voice_analysis_summary),
+                    self._evaluate_job_fit(req, transcript, rag_context),
+                    self._coach_answers(req, rag_context),
+                )
             )
             if self_intro_item is not None:
                 result.panel_breakdown.append(self_intro_item)
@@ -141,6 +150,7 @@ async def handle(self, message: AbstractIncomingMessage) -> None:
                 improvement_keywords=result.improvement_keywords,
                 study_plan=result.study_plan,
                 panel_breakdown=result.panel_breakdown,
+                answer_coaching=answer_coaching,
                 report_s3_key=None,
             )
 
@@ -227,6 +237,56 @@ async def _evaluate_job_fit(
             ),
         ]
 
+    async def _coach_answers(
+        self, req: GenerateFeedbackRequest, rag_context: str
+    ) -> list[AnswerCoachingItem]:
+        """자기소개 제외 답변마다 모범 답안·리라이트·코칭을 병렬 생성 → 메시지별 복기 리스트."""
+        if self._answer_coach is None:
+            return []
+        pairs = _collect_coachable_pairs(req.messages)
+        if not pairs:
+            return []
+        if len(pairs) > self._coaching_max_answers:
+            log.info(
+                "feedback.coaching.capped",
+                session_id=req.session_id,
+                total=len(pairs),
+                cap=self._coaching_max_answers,
+            )
+            pairs = pairs[: self._coaching_max_answers]
+        target_role = _coaching_target_role(req)
+
+        async def _one(
+            question: FeedbackMessageItem, answer: FeedbackMessageItem
+        ) -> AnswerCoachingItem | None:
+            try:
+                res = await self._answer_coach.coach(
+                    job_category=req.job_category,
+                    mode=req.mode,
+                    target_role=target_role,
+                    question=question.content,
+                    expected_signal=question.expected_signal or "",
+                    answer=answer.content,
+                    rag_context=rag_context,
+                )
+            except Exception as exc:  # noqa: BLE001
+                log.warning(
+                    "feedback.coaching.failed",
+                    error=str(exc),
+                    session_id=req.session_id,
+                    message_id=answer.id,
+                )
+                return None
+            return AnswerCoachingItem(
+                message_id=answer.id,
+                model_answer=res.model_answer,
+                answer_rewrite=res.answer_rewrite,
+                coaching_comment=res.coaching_comment,
+            )
+
+        items = await asyncio.gather(*(_one(q, a) for q, a in pairs))
+        return [it for it in items if it is not None]
+
     async def _build_rag_context(self, req: GenerateFeedbackRequest) -> str:
         if not self._embedder or not req.context_document_ids:
             return "(none)"
@@ -256,6 +316,36 @@ async def _build_rag_context(self, req: GenerateFeedbackRequest) -> str:
         )
 
 
+def _collect_coachable_pairs(
+    messages: list[FeedbackMessageItem],
+) -> list[tuple[FeedbackMessageItem, FeedbackMessageItem]]:
+    """복기 대상 (질문, 답변) 쌍. 자기소개 질문에 대한 답변과 빈 답변은 제외, 시퀀스 순."""
+    by_id = {m.id: m for m in messages}
+    pairs: list[tuple[FeedbackMessageItem, FeedbackMessageItem]] = []
+    for m in messages:
+        if m.role != "INTERVIEWEE" or not (m.content or "").strip():
+            continue
+        question = by_id.get(m.parent_message_id) if m.parent_message_id else None
+        if question is None or question.role != "INTERVIEWER":
+            continue
+        if (question.category or "") == _SELF_INTRO_CATEGORY:
+            continue  # 자기소개는 첫인상 평가가 커버
+        pairs.append((question, m))
+    return pairs
+
+
+def _coaching_target_role(req: GenerateFeedbackRequest) -> str:
+    """직무 맞춤 모드일 때만 코칭 프롬프트에 실을 회사/JD 발췌. 그 외 빈 문자열."""
+    if (req.mode or "") != _JOB_TAILORED_MODE:
+        return ""
+    jd = (req.target_job_description or "").strip()
+    if not jd:
+        return ""
+    company = (req.target_company_name or "").strip()
+    head = f"지원 회사: {company}\n" if company else ""
+    return f"{head}채용공고(JD) 발췌: {jd[:800]}"
+
+
 def _to_panel_item(
     label: str, dimension: str, ev: EvaluatorResult
 ) -> PanelBreakdownItem:
diff --git a/ai/src/ai_server/messaging/runner.py b/ai/src/ai_server/messaging/runner.py
index 102886ba..c62610f8 100644
--- a/ai/src/ai_server/messaging/runner.py
+++ b/ai/src/ai_server/messaging/runner.py
@@ -20,9 +20,11 @@
     build_streaming_followup_generator,
 )
 from ai_server.chain.feedback_generation_chain import (
+    LlmAnswerCoach,
     LlmJobFitEvaluator,
     LlmSelfIntroEvaluator,
     PanelFeedbackGenerator,
+    build_answer_coaching_chain,
     build_feedback_synthesis_chain,
     build_job_fit_evaluation_chain,
     build_panel_evaluator_chain,
@@ -239,6 +241,10 @@ def __init__(self, settings: Settings) -> None:
             job_fit_evaluator=LlmJobFitEvaluator(
                 build_job_fit_evaluation_chain(settings, core_client=core_client)
             ),
+            # 질문별 복기(Flash, 답변 수만큼 병렬). 모범 답안+리라이트+한 줄 코칭.
+            answer_coach=LlmAnswerCoach(
+                build_answer_coaching_chain(settings, core_client=core_client)
+            ),
         )
 
         # 음성 답변 STT + 분석 (Phase 2)
diff --git a/ai/src/ai_server/model/messages/feedback.py b/ai/src/ai_server/model/messages/feedback.py
index 5cd3dcea..0b9ca553 100644
--- a/ai/src/ai_server/model/messages/feedback.py
+++ b/ai/src/ai_server/model/messages/feedback.py
@@ -85,6 +85,17 @@ class PanelBreakdownItem(BaseModel):
     score_rationale: str | None = None
 
 
+class AnswerCoachingItem(BaseModel):
+    """질문별 복기 1건 (AI → Core). 해당 답변(INTERVIEWEE) 메시지에 기록된다."""
+
+    model_config = camel_config()
+
+    message_id: int
+    model_answer: str | None = None
+    answer_rewrite: str | None = None
+    coaching_comment: str | None = None
+
+
 class FeedbackCallbackPayload(BaseModel):
     """AI → Core 종합 피드백. 점수는 0~100 (NULL 허용)."""
 
@@ -102,4 +113,6 @@ class FeedbackCallbackPayload(BaseModel):
     study_plan: list[str] = Field(default_factory=list)
     # 평가위원별 분해(패널). 비어 있으면 단일/레거시 경로.
     panel_breakdown: list[PanelBreakdownItem] = Field(default_factory=list)
+    # 질문별 복기(답변 메시지별 모범 답안·리라이트·코칭). 비면 복기 없음.
+    answer_coaching: list[AnswerCoachingItem] = Field(default_factory=list)
     report_s3_key: str | None = None
diff --git a/ai/tests/test_feedback_consumer.py b/ai/tests/test_feedback_consumer.py
index 5d58aeef..fe9ad752 100644
--- a/ai/tests/test_feedback_consumer.py
+++ b/ai/tests/test_feedback_consumer.py
@@ -6,9 +6,11 @@
 import pytest
 
 from ai_server.chain.feedback_generation_chain import (
+    CoachingResult,
     EvaluatorResult,
     FeedbackResult,
     JobFitResult,
+    LlmAnswerCoach,
     LlmFeedbackGenerator,
     LlmJobFitEvaluator,
     LlmSelfIntroEvaluator,
@@ -672,6 +674,131 @@ async def ainvoke(self, value):
     assert result.understanding.score == 55.0
 
 
+def _answer_coach():
+    c = MagicMock()
+    c.coach = AsyncMock(
+        return_value=CoachingResult(
+            model_answer="격리수준별 이상현상과 트레이드오프를 설명…",
+            answer_rewrite="두괄식으로 핵심 먼저, 그다음 근거…",
+            coaching_comment="결론을 먼저 말하고 근거로 받치세요.",
+        )
+    )
+    return c
+
+
+@pytest.mark.asyncio
+async def test_consumer_appends_answer_coaching_excluding_self_intro():
+    generator = _generator()
+    publisher = MagicMock()
+    publisher.publish = AsyncMock()
+    coach = _answer_coach()
+
+    consumer = FeedbackConsumer(
+        generator=generator,
+        publisher=publisher,
+        idempotency=LruIdempotencyStore(max_size=10),
+        callback_routing_key="callback.feedback",
+        core_client=MagicMock(),
+        embedder=None,
+        answer_coach=coach,
+    )
+    # _self_intro_envelope: 자기소개 Q(200)/A(201) + ACID Q(202)/A(203).
+    await consumer.handle(_StubMessage(_self_intro_envelope()))
+
+    # 자기소개 답변(201)은 코칭 제외, ACID 답변(203)만 코칭.
+    assert coach.coach.await_count == 1
+    assert "ACID" in coach.coach.await_args.kwargs["question"]
+
+    payload: FeedbackCallbackPayload = publisher.publish.await_args.kwargs["payload"]
+    assert len(payload.answer_coaching) == 1
+    item = payload.answer_coaching[0]
+    assert item.message_id == 203
+    assert item.model_answer.startswith("격리수준")
+    assert all(c.message_id != 201 for c in payload.answer_coaching)
+
+
+def test_collect_coachable_pairs_excludes_self_intro_and_empty():
+    from ai_server.messaging.consumers.feedback_consumer import _collect_coachable_pairs
+
+    msgs = [
+        FeedbackMessageItem(
+            id=1,
+            sequence_number=1,
+            role="INTERVIEWER",
+            content="자기소개?",
+            category="SELF_INTRODUCTION",
+        ),
+        FeedbackMessageItem(
+            id=2,
+            sequence_number=2,
+            role="INTERVIEWEE",
+            content="소개",
+            parent_message_id=1,
+        ),
+        FeedbackMessageItem(
+            id=3,
+            sequence_number=3,
+            role="INTERVIEWER",
+            content="ACID?",
+            category="CS_FUNDAMENTAL",
+        ),
+        FeedbackMessageItem(
+            id=4,
+            sequence_number=4,
+            role="INTERVIEWEE",
+            content="원자성…",
+            parent_message_id=3,
+        ),
+        # 빈 답변은 제외
+        FeedbackMessageItem(
+            id=5,
+            sequence_number=5,
+            role="INTERVIEWER",
+            content="Q2?",
+            category="TECH_CHOICE",
+        ),
+        FeedbackMessageItem(
+            id=6,
+            sequence_number=6,
+            role="INTERVIEWEE",
+            content="   ",
+            parent_message_id=5,
+        ),
+    ]
+    pairs = _collect_coachable_pairs(msgs)
+    assert [(q.id, a.id) for q, a in pairs] == [(3, 4)]
+
+
+@pytest.mark.asyncio
+async def test_answer_coach_forwards_inputs_to_chain():
+    class _FakeChain:
+        def __init__(self):
+            self.input = None
+
+        async def ainvoke(self, value):
+            self.input = value
+            return CoachingResult(
+                model_answer="m", answer_rewrite="r", coaching_comment="c"
+            )
+
+    chain = _FakeChain()
+    coach = LlmAnswerCoach(chain)
+
+    await coach.coach(
+        job_category="BACKEND",
+        mode="TECHNICAL",
+        target_role="",
+        question="트랜잭션 격리수준?",
+        expected_signal="격리수준·이상현상",
+        answer="음… 잘 모르겠습니다",
+        rag_context="resume chunk",
+    )
+
+    assert chain.input["question"] == "트랜잭션 격리수준?"
+    assert chain.input["expected_signal"] == "격리수준·이상현상"
+    assert chain.input["answer"] == "음… 잘 모르겠습니다"
+
+
 def test_build_transcript_annotates_interviewee_evaluation():
     from ai_server.messaging.consumers.feedback_consumer import _build_transcript
     from ai_server.model.messages.feedback import FeedbackMessageItem, MessageEvaluation
diff --git a/backend/CLAUDE.md b/backend/CLAUDE.md
index 603d9f4c..dc4adc66 100644
--- a/backend/CLAUDE.md
+++ b/backend/CLAUDE.md
@@ -359,6 +359,11 @@ docker compose up -d
 - ArchUnit 룰 적용 (의존 방향 · 순환 차단 · `@Transactional` application 한정 · entity는 domain 패키지)
 - 면접 도메인 (US-13~20) 본 구현: 세션 CRUD/start/end/interrupt, generate.questions 발행,
   callback.questions(POOL/FOLLOWUP) 수신, 자동 종료
+- **질문별 복기 본 구현**: V19 로 `interview_messages` 에 `model_answer`/`answer_rewrite`/`coaching_comment`
+  추가. AI 가 `callback.feedback.answerCoaching[{messageId,…}]` 로 답변별 모범 답안·리라이트·코칭을 보내면
+  `FeedbackCallbackService` 가 각 메시지에 `recordCoaching` 기록. `MessageResult`/`MessageResponse` 가 답변
+  평가 점수 + 복기를 노출하되 **종료 세션 조회에서만**(`expectedSignal` 과 동일 게이팅). 프론트는 답변 버블
+  아래 '복기' 아코디언으로 표시.
 - **직무 맞춤 면접 모드(JOB_TAILORED) 본 구현**: `SessionMode.JOB_TAILORED` 추가(V18 — mode CHECK 갱신 +
   `target_company_name`/`target_job_description` 컬럼). 이 모드는 회사명+채용공고(JD)를 받아(JD 필수,
   `SessionService` 검증 `SESSION_JD_REQUIRED`) `InterviewSession.assignTargetRole` 로 보관. JD 는
diff --git a/backend/openapi.json b/backend/openapi.json
index 37ee10b0..7bb832c4 100644
--- a/backend/openapi.json
+++ b/backend/openapi.json
@@ -2804,6 +2804,30 @@
           },
           "expectedSignal" : {
             "type" : "string"
+          },
+          "answerSpecificity" : {
+            "type" : "number",
+            "format" : "double"
+          },
+          "answerLogic" : {
+            "type" : "number",
+            "format" : "double"
+          },
+          "answerStructure" : {
+            "type" : "string"
+          },
+          "answerCorrectness" : {
+            "type" : "number",
+            "format" : "double"
+          },
+          "modelAnswer" : {
+            "type" : "string"
+          },
+          "answerRewrite" : {
+            "type" : "string"
+          },
+          "coachingComment" : {
+            "type" : "string"
           }
         }
       },
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/FeedbackCallbackService.java b/backend/src/main/java/com/stackup/stackup/session/application/FeedbackCallbackService.java
index 45b31cd3..fe4d0fe7 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/FeedbackCallbackService.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/FeedbackCallbackService.java
@@ -7,8 +7,11 @@
 import com.stackup.stackup.common.messaging.RealtimeNotifyEvent;
 import com.stackup.stackup.common.sse.SseEventType;
 import org.springframework.context.ApplicationEventPublisher;
+import com.stackup.stackup.session.application.dto.AnswerCoachingItem;
 import com.stackup.stackup.session.application.dto.FeedbackCallbackEnvelope;
 import com.stackup.stackup.session.application.dto.FeedbackCallbackPayload;
+import com.stackup.stackup.session.domain.InterviewMessage;
+import com.stackup.stackup.session.domain.InterviewMessageRepository;
 import com.stackup.stackup.session.domain.InterviewSession;
 import com.stackup.stackup.session.domain.InterviewSessionRepository;
 import com.stackup.stackup.session.domain.SessionFeedback;
@@ -33,6 +36,7 @@ public class FeedbackCallbackService {
 
     private final InterviewSessionRepository sessionRepository;
     private final SessionFeedbackRepository feedbackRepository;
+    private final InterviewMessageRepository messageRepository;
     private final ProcessedMessageRepository processedMessageRepository;
     private final ApplicationEventPublisher events;
 
@@ -87,6 +91,8 @@ public void apply(FeedbackCallbackEnvelope envelope) {
             return;
         }
 
+        applyAnswerCoaching(sessionId, payload.answerCoaching());
+
         events.publishEvent(RealtimeNotifyEvent.session(sessionId, SseEventType.FEEDBACK_READY,
             new SessionFeedbackNotice(sessionId, feedback.getId())));
         events.publishEvent(RealtimeNotifyEvent.user(session.getUser().getId(), SseEventType.FEEDBACK_READY,
@@ -96,6 +102,24 @@ public void apply(FeedbackCallbackEnvelope envelope) {
         log.info("callback.feedback processed. sessionId={}, feedbackId={}", sessionId, feedback.getId());
     }
 
+    // 답변별 복기를 각 INTERVIEWEE 메시지에 기록. 메시지가 다른 세션이면 방어적으로 skip.
+    private void applyAnswerCoaching(Long sessionId, java.util.List<AnswerCoachingItem> items) {
+        if (items == null || items.isEmpty()) {
+            return;
+        }
+        for (AnswerCoachingItem item : items) {
+            if (item == null || item.messageId() == null) {
+                continue;
+            }
+            InterviewMessage message = messageRepository.findById(item.messageId()).orElse(null);
+            if (message == null || !sessionId.equals(message.getSession().getId())) {
+                continue;
+            }
+            message.recordCoaching(item.modelAnswer(), item.answerRewrite(), item.coachingComment());
+            messageRepository.save(message);
+        }
+    }
+
     public record SessionFeedbackNotice(Long sessionId, Long feedbackId) {
     }
 
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java b/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
index c65ff62b..50fc4d7b 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
@@ -51,12 +51,11 @@ public List<MessageResult> list(Long userId, Long sessionId) {
 
     // 재생용 presigned URL 동봉: 질문 TTS(SUCCEEDED) + 음성 답변 원본.
     // presign 실패가 메시지 조회 전체를 깨뜨리지 않도록 개별 try/catch.
-    private MessageResult toResultWithAudioUrls(InterviewMessage m, boolean revealExpectedSignal) {
+    private MessageResult toResultWithAudioUrls(InterviewMessage m, boolean revealInsights) {
         String ttsUrl = m.getTtsStatus() == TtsStatus.SUCCEEDED
             ? presign(m.getTtsAudioPath()) : null;
         String audioUrl = presign(m.getAudioFilePath());
-        String expectedSignal = revealExpectedSignal ? m.getExpectedSignal() : null;
-        return MessageResult.of(m, ttsUrl, audioUrl, expectedSignal);
+        return MessageResult.of(m, ttsUrl, audioUrl, revealInsights);
     }
 
     private String presign(String key) {
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/AnswerCoachingItem.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/AnswerCoachingItem.java
new file mode 100644
index 00000000..9dfae7fd
--- /dev/null
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/AnswerCoachingItem.java
@@ -0,0 +1,10 @@
+package com.stackup.stackup.session.application.dto;
+
+// callback.feedback 의 답변별 복기 1건 (AI → Core). 해당 답변(INTERVIEWEE) 메시지에 기록된다.
+public record AnswerCoachingItem(
+    Long messageId,
+    String modelAnswer,
+    String answerRewrite,
+    String coachingComment
+) {
+}
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/FeedbackCallbackPayload.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/FeedbackCallbackPayload.java
index d9dad1d1..fe7dc81f 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/dto/FeedbackCallbackPayload.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/FeedbackCallbackPayload.java
@@ -16,6 +16,8 @@ public record FeedbackCallbackPayload(
     List<String> improvementKeywords,
     List<String> studyPlan,
     List<PanelBreakdownItem> panelBreakdown,
+    // 질문별 복기 (답변 메시지별 모범 답안·리라이트·코칭). 비면 복기 없음.
+    List<AnswerCoachingItem> answerCoaching,
     String reportS3Key
 ) {
 }
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
index 26b3fc97..309c9c6b 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
@@ -26,18 +26,27 @@ public record MessageResult(
     String audioFileUrl,
     // 질문이 기대한 핵심(평가 관점). 라이브 중엔 null(정답 유출 방지),
     // 종료된 세션 조회에서만 채워 피드백 학습용으로 노출.
-    String expectedSignal
+    String expectedSignal,
+    // 답변별 복기 — 종료 세션 조회에서만 노출(expectedSignal 과 동일 게이팅). 라이브 중엔 모두 null.
+    Double answerSpecificity,
+    Double answerLogic,
+    String answerStructure,
+    Double answerCorrectness,
+    String modelAnswer,
+    String answerRewrite,
+    String coachingComment
 ) {
     public static MessageResult of(InterviewMessage m) {
-        return of(m, null, null, null);
+        return of(m, null, null, false);
     }
 
     public static MessageResult of(InterviewMessage m, String ttsAudioUrl, String audioFileUrl) {
-        return of(m, ttsAudioUrl, audioFileUrl, null);
+        return of(m, ttsAudioUrl, audioFileUrl, false);
     }
 
+    // revealInsights=true 면 답변 평가·복기·expectedSignal 노출(종료 세션). 라이브/대기 중엔 false.
     public static MessageResult of(
-        InterviewMessage m, String ttsAudioUrl, String audioFileUrl, String expectedSignal) {
+        InterviewMessage m, String ttsAudioUrl, String audioFileUrl, boolean revealInsights) {
         return new MessageResult(
             m.getId(),
             m.getSession().getId(),
@@ -55,7 +64,14 @@ public static MessageResult of(
             m.getTargetEvidence(),
             ttsAudioUrl,
             audioFileUrl,
-            expectedSignal
+            revealInsights ? m.getExpectedSignal() : null,
+            revealInsights ? m.getAnswerSpecificity() : null,
+            revealInsights ? m.getAnswerLogic() : null,
+            revealInsights ? m.getAnswerStructure() : null,
+            revealInsights ? m.getAnswerCorrectness() : null,
+            revealInsights ? m.getModelAnswer() : null,
+            revealInsights ? m.getAnswerRewrite() : null,
+            revealInsights ? m.getCoachingComment() : null
         );
     }
 }
diff --git a/backend/src/main/java/com/stackup/stackup/session/domain/InterviewMessage.java b/backend/src/main/java/com/stackup/stackup/session/domain/InterviewMessage.java
index 1df8c4e1..3e7972c5 100644
--- a/backend/src/main/java/com/stackup/stackup/session/domain/InterviewMessage.java
+++ b/backend/src/main/java/com/stackup/stackup/session/domain/InterviewMessage.java
@@ -109,6 +109,16 @@ public class InterviewMessage extends BaseTimeEntity {
     @Column(name = "answer_correctness")
     private Double answerCorrectness;
 
+    // 질문별 복기 (INTERVIEWEE 답변에만 채워짐). 피드백 생성 시 AI가 만든 모범 답안·리라이트·코칭.
+    @Column(name = "model_answer", columnDefinition = "text")
+    private String modelAnswer;
+
+    @Column(name = "answer_rewrite", columnDefinition = "text")
+    private String answerRewrite;
+
+    @Column(name = "coaching_comment", columnDefinition = "text")
+    private String coachingComment;
+
     // 부연(질문 재설명) 메시지 여부. true 면 총 질문 수·꼬리 깊이에 카운트하지 않는다.
     @Column(nullable = false)
     private boolean clarification = false;
@@ -218,6 +228,13 @@ public void recordAnswerEvaluation(Double specificity, Double logic,
         this.answerCorrectness = correctness;
     }
 
+    // 피드백 생성 시 답변별 복기(모범 답안·리라이트·코칭)를 이 답변 메시지에 기록.
+    public void recordCoaching(String modelAnswer, String answerRewrite, String coachingComment) {
+        this.modelAnswer = modelAnswer;
+        this.answerRewrite = answerRewrite;
+        this.coachingComment = coachingComment;
+    }
+
     public void attachAudio(String audioFilePath) {
         this.audioFilePath = audioFilePath;
     }
diff --git a/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java b/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
index a8272f96..8ec27946 100644
--- a/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
+++ b/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
@@ -25,7 +25,15 @@ public record MessageResponse(
     String ttsAudioUrl,
     String audioFileUrl,
     // 질문이 기대한 핵심(평가 관점). 종료된 세션에서만 채워짐(라이브 중엔 null — 정답 유출 방지).
-    String expectedSignal
+    String expectedSignal,
+    // 답변별 복기(종료 세션에서만). 답변 평가 점수 + 모범 답안 + 리라이트 + 한 줄 코칭.
+    Double answerSpecificity,
+    Double answerLogic,
+    String answerStructure,
+    Double answerCorrectness,
+    String modelAnswer,
+    String answerRewrite,
+    String coachingComment
 ) {
     public static MessageResponse from(MessageResult r) {
         return new MessageResponse(
@@ -45,7 +53,14 @@ public static MessageResponse from(MessageResult r) {
             r.targetEvidence(),
             r.ttsAudioUrl(),
             r.audioFileUrl(),
-            r.expectedSignal()
+            r.expectedSignal(),
+            r.answerSpecificity(),
+            r.answerLogic(),
+            r.answerStructure(),
+            r.answerCorrectness(),
+            r.modelAnswer(),
+            r.answerRewrite(),
+            r.coachingComment()
         );
     }
 }
diff --git a/backend/src/main/resources/db/migration/V19__add_answer_coaching_to_interview_messages.sql b/backend/src/main/resources/db/migration/V19__add_answer_coaching_to_interview_messages.sql
new file mode 100644
index 00000000..ede0a578
--- /dev/null
+++ b/backend/src/main/resources/db/migration/V19__add_answer_coaching_to_interview_messages.sql
@@ -0,0 +1,7 @@
+-- 질문별 복기: 답변마다 AI가 생성한 모범 답안 + 내 답변 리라이트 + 한 줄 코칭.
+-- 피드백 생성(callback.feedback) 시 답변(INTERVIEWEE) 메시지에 기록되고, 종료 세션 조회에서만 노출된다.
+
+ALTER TABLE interview_messages
+    ADD COLUMN model_answer     TEXT,
+    ADD COLUMN answer_rewrite   TEXT,
+    ADD COLUMN coaching_comment TEXT;
diff --git a/backend/src/test/java/com/stackup/stackup/session/application/FeedbackCallbackServiceTest.java b/backend/src/test/java/com/stackup/stackup/session/application/FeedbackCallbackServiceTest.java
index cef5a58f..0539e7e1 100644
--- a/backend/src/test/java/com/stackup/stackup/session/application/FeedbackCallbackServiceTest.java
+++ b/backend/src/test/java/com/stackup/stackup/session/application/FeedbackCallbackServiceTest.java
@@ -10,9 +10,12 @@
 import com.stackup.stackup.common.messaging.RealtimeNotifyEvent;
 import com.stackup.stackup.common.messaging.domain.ProcessedMessageRepository;
 import com.stackup.stackup.common.sse.SseEventType;
+import com.stackup.stackup.session.application.dto.AnswerCoachingItem;
 import com.stackup.stackup.session.application.dto.FeedbackCallbackEnvelope;
 import com.stackup.stackup.session.application.dto.FeedbackCallbackPayload;
 import com.stackup.stackup.session.application.dto.PanelBreakdownItem;
+import com.stackup.stackup.session.domain.InterviewMessage;
+import com.stackup.stackup.session.domain.InterviewMessageRepository;
 import com.stackup.stackup.session.domain.InterviewSession;
 import com.stackup.stackup.session.domain.InterviewSessionRepository;
 import com.stackup.stackup.session.domain.JobCategory;
@@ -36,6 +39,7 @@ class FeedbackCallbackServiceTest {
 
     @Mock InterviewSessionRepository sessionRepository;
     @Mock SessionFeedbackRepository feedbackRepository;
+    @Mock InterviewMessageRepository messageRepository;
     @Mock ProcessedMessageRepository processedMessageRepository;
     @Mock ApplicationEventPublisher events;
     @InjectMocks FeedbackCallbackService service;
@@ -49,7 +53,7 @@ void apply_insertsFeedbackAndPushesSse() {
                 List.of("Redis 분산 락 직접 구현"),
                 List.of(new PanelBreakdownItem("기술", "기술 정확도·깊이", 80.0,
                     "설계 깊이", "테스트 부족", "상세 평가 문단", "근거")),
-                null));
+                List.of(), null));
 
         when(processedMessageRepository.existsById("fb-1")).thenReturn(false);
         when(feedbackRepository.existsBySession_Id(50L)).thenReturn(false);
@@ -79,10 +83,36 @@ void apply_insertsFeedbackAndPushesSse() {
         });
     }
 
+    @Test
+    void apply_writesAnswerCoachingToMessages() {
+        InterviewSession session = sessionFixture(50L);
+        InterviewMessage answer = InterviewMessage.interviewee(session, 2, "내 답변", null, null);
+        ReflectionTestUtils.setField(answer, "id", 600L);
+
+        FeedbackCallbackEnvelope env = envelope(50L, "fb-coach",
+            new FeedbackCallbackPayload(50L, 80.0, null, null, null, null, null,
+                List.of(), List.of(), List.of(),
+                List.of(new AnswerCoachingItem(600L, "모범 답안", "리라이트", "두괄식으로")),
+                null));
+
+        when(processedMessageRepository.existsById("fb-coach")).thenReturn(false);
+        when(feedbackRepository.existsBySession_Id(50L)).thenReturn(false);
+        when(sessionRepository.findById(50L)).thenReturn(Optional.of(session));
+        when(feedbackRepository.save(any(SessionFeedback.class))).thenAnswer(inv -> inv.getArgument(0));
+        when(messageRepository.findById(600L)).thenReturn(Optional.of(answer));
+
+        service.apply(env);
+
+        assertThat(answer.getModelAnswer()).isEqualTo("모범 답안");
+        assertThat(answer.getAnswerRewrite()).isEqualTo("리라이트");
+        assertThat(answer.getCoachingComment()).isEqualTo("두괄식으로");
+        verify(messageRepository).save(answer);
+    }
+
     @Test
     void apply_skipsWhenDuplicateMessage() {
         FeedbackCallbackEnvelope env = envelope(50L, "dup",
-            new FeedbackCallbackPayload(50L, 80.0, null, null, null, null, null, List.of(), List.of(), List.of(), null));
+            new FeedbackCallbackPayload(50L, 80.0, null, null, null, null, null, List.of(), List.of(), List.of(), List.of(), null));
         when(processedMessageRepository.existsById("dup")).thenReturn(true);
 
         service.apply(env);
@@ -94,7 +124,7 @@ void apply_skipsWhenDuplicateMessage() {
     void apply_skipsWhenFeedbackAlreadyExists() {
         InterviewSession session = sessionFixture(50L);
         FeedbackCallbackEnvelope env = envelope(50L, "fb-2",
-            new FeedbackCallbackPayload(50L, 80.0, null, null, null, null, null, List.of(), List.of(), List.of(), null));
+            new FeedbackCallbackPayload(50L, 80.0, null, null, null, null, null, List.of(), List.of(), List.of(), List.of(), null));
         when(processedMessageRepository.existsById("fb-2")).thenReturn(false);
         when(sessionRepository.findById(50L)).thenReturn(Optional.of(session));
         when(feedbackRepository.existsBySession_Id(50L)).thenReturn(true);
diff --git a/docs/database.md b/docs/database.md
index d89c99c8..a2d7ddc0 100644
--- a/docs/database.md
+++ b/docs/database.md
@@ -193,6 +193,10 @@ CREATE TABLE interview_messages (
     UNIQUE (session_id, sequence_number),
     CHECK (content IS NOT NULL OR audio_file_path IS NOT NULL)
 );
+-- interview_messages 는 이후 마이그레이션으로 컬럼 추가:
+--   tts_* (V7), category/target_evidence/expected_signal (V9), answer_* 평가 4종 (V10),
+--   clarification (V13), 그리고 질문별 복기 3종 (V19):
+--     model_answer TEXT, answer_rewrite TEXT, coaching_comment TEXT  -- 답변(INTERVIEWEE)에만, 종료 세션 조회에서만 노출
 
 -- 10. message_voice_analyses
 CREATE TABLE message_voice_analyses (
diff --git a/docs/messaging.md b/docs/messaging.md
index 7c3e93f9..1defbfc0 100644
--- a/docs/messaging.md
+++ b/docs/messaging.md
@@ -368,6 +368,8 @@
 
 ### 5.11 `callback.feedback`
 
+> `answerCoaching[]` 은 질문별 복기 — 답변(INTERVIEWEE) 메시지별 모범 답안·리라이트·한 줄 코칭(자기소개 제외).
+> Core 가 각 `messageId` 의 메시지에 기록하고 종료 세션 조회에서만 노출한다.
 > `panelBreakdown[]` 에 평가위원별 항목이 담긴다. 자기소개가 있던 세션은 **`evaluator="첫인상"`**,
 > 직무 맞춤 모드는 **`evaluator="직무 적합도"`(역량 매칭)** + **`evaluator="직무 이해도"`(직무 이해·동기)**
 > 항목이 추가로 포함된다 — 모두 **종합 점수(overallScore) 집계에서 제외**된 별도 정성 평가다(메인
@@ -386,6 +388,9 @@
     "weaknessesSummary": "...",
     "improvementKeywords": ["JPA 영속성 컨텍스트", "TCP 3-way handshake"],
     "studyPlan": ["..."],
+    "answerCoaching": [
+      { "messageId": 203, "modelAnswer": "이 질문에 강한 답변 예시…", "answerRewrite": "내 답변을 이렇게 고치면…", "coachingComment": "결론을 먼저 말하세요." }
+    ],
     "panelBreakdown": [
       { "evaluator": "백엔드", "dimension": "기술 정확도·깊이", "score": 80.0, "detail": "...", "scoreRationale": "..." },
       { "evaluator": "첫인상", "dimension": "자기소개 전달력·구성·직무적합성", "score": 78.0, "detail": "...", "scoreRationale": "..." }
diff --git a/frontend/src/features/interview/ui/InterviewTranscript.tsx b/frontend/src/features/interview/ui/InterviewTranscript.tsx
index 9a1a10e2..383efb2c 100644
--- a/frontend/src/features/interview/ui/InterviewTranscript.tsx
+++ b/frontend/src/features/interview/ui/InterviewTranscript.tsx
@@ -3,6 +3,7 @@ import { Spinner } from '@/shared/ui/Spinner'
 import { useSessionMessages } from '../model/useSessionMessages'
 import { QuestionBubble } from './live/QuestionBubble'
 import { AnswerBubble } from './live/AnswerBubble'
+import { AnswerCoachingAccordion } from './live/AnswerCoachingAccordion'
 
 // 종료된 세션의 질문/답변 원문(읽기 전용). 라이브 버블을 재사용해
 // 음성 질문 재생(TTS)·음성 답변 재생도 그대로 노출한다.
@@ -30,7 +31,10 @@ export function InterviewTranscript({ sessionId }: { sessionId: number }) {
           isQuestion(m) ? (
             <QuestionBubble key={m.id} message={m} />
           ) : (
-            <AnswerBubble key={m.id} message={m} />
+            <div key={m.id} className="flex flex-col items-end gap-1">
+              <AnswerBubble message={m} />
+              <AnswerCoachingAccordion message={m} />
+            </div>
           ),
         )}
       </div>
diff --git a/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx b/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx
new file mode 100644
index 00000000..4464fcd1
--- /dev/null
+++ b/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx
@@ -0,0 +1,85 @@
+import { useState } from 'react'
+import type { Message } from '@/domain/session'
+
+// 답변 구조 채점값(FULL_STAR/PARTIAL_STAR/NONE) → 한국어 라벨.
+const STRUCTURE_LABEL: Record<string, string> = {
+  FULL_STAR: '구조 양호',
+  PARTIAL_STAR: '구조 보통',
+  NONE: '구조 미흡',
+}
+
+function score5(v?: number | null): string | null {
+  return typeof v === 'number' ? `${v}/5` : null
+}
+
+// 종료된 세션의 답변 아래 붙는 '복기' 아코디언: 평가 점수 → 한 줄 코칭 → 모범 답안 → 내 답변 리라이트.
+// 코칭 데이터가 없으면(라이브 중·자기소개 답변 등) 렌더하지 않는다.
+export function AnswerCoachingAccordion({ message }: { message: Message }) {
+  const [open, setOpen] = useState(false)
+  const hasCoaching = Boolean(
+    message.modelAnswer || message.answerRewrite || message.coachingComment,
+  )
+  if (!hasCoaching) return null
+
+  const spec = score5(message.answerSpecificity)
+  const logic = score5(message.answerLogic)
+  const correctness = score5(message.answerCorrectness)
+  const evals = [
+    spec && `구체성 ${spec}`,
+    logic && `논리 ${logic}`,
+    correctness && `정답성 ${correctness}`,
+    message.answerStructure
+      ? (STRUCTURE_LABEL[message.answerStructure] ?? message.answerStructure)
+      : null,
+  ].filter(Boolean) as string[]
+
+  return (
+    <div className="w-full max-w-[80%] self-end">
+      <button
+        type="button"
+        onClick={() => setOpen((o) => !o)}
+        className="text-caption text-fg-muted transition-colors hover:text-fg"
+      >
+        {open ? '복기 닫기 ▴' : '복기 보기 ▾'}
+      </button>
+      {open && (
+        <div className="mt-1.5 flex flex-col gap-3 rounded-lg border border-border bg-surface px-4 py-3 text-left">
+          {evals.length > 0 && (
+            <div className="flex flex-wrap gap-1.5">
+              {evals.map((e) => (
+                <span
+                  key={e}
+                  className="rounded-pill bg-surface-raised px-2 py-0.5 text-caption text-fg-muted"
+                >
+                  {e}
+                </span>
+              ))}
+            </div>
+          )}
+          {message.coachingComment && (
+            <p className="text-body text-fg">
+              <span className="text-primary">코칭 · </span>
+              {message.coachingComment}
+            </p>
+          )}
+          {message.modelAnswer && (
+            <div className="flex flex-col gap-1">
+              <span className="text-caption text-fg-subtle">모범 답안</span>
+              <p className="whitespace-pre-wrap text-body text-fg-muted">
+                {message.modelAnswer}
+              </p>
+            </div>
+          )}
+          {message.answerRewrite && (
+            <div className="flex flex-col gap-1">
+              <span className="text-caption text-fg-subtle">내 답변, 이렇게 고치면</span>
+              <p className="whitespace-pre-wrap text-body text-fg-muted">
+                {message.answerRewrite}
+              </p>
+            </div>
+          )}
+        </div>
+      )}
+    </div>
+  )
+}
diff --git a/frontend/src/shared/api/generated.ts b/frontend/src/shared/api/generated.ts
index 05b8aa57..917f940b 100644
--- a/frontend/src/shared/api/generated.ts
+++ b/frontend/src/shared/api/generated.ts
@@ -900,6 +900,16 @@ export interface components {
             ttsAudioUrl?: string;
             audioFileUrl?: string;
             expectedSignal?: string;
+            /** Format: double */
+            answerSpecificity?: number;
+            /** Format: double */
+            answerLogic?: number;
+            answerStructure?: string;
+            /** Format: double */
+            answerCorrectness?: number;
+            modelAnswer?: string;
+            answerRewrite?: string;
+            coachingComment?: string;
         };
         VoiceStreamBeginResponse: {
             /** Format: int64 */