优化住流程和添加复刻视频页面

core/backend/apps/ai/script_agent.py

@@ -68,6 +68,51 @@ STRUCTURE_MIN_DURATION: dict[str, int] = {
     "pain": 15, "contrast": 15, "review": 30, "scene": 30, "promo": 15, "knowledge": 30,
 }
 
+# 用户在设定卡里选的是「怎么说服」，不是一个仅供展示的标签。这里把每种结构收成
+# 可直接塞进任务上下文的验收合同；模型不得拿痛点结构去代替知识分享、拿种草去代替测评。
+STRUCTURE_CONTRACTS: dict[str, str] = {
+    "pain": (
+        "【痛点解决·结构合同】必须按顺序完成：具体问题出现 → 这个问题造成的不便 → 商品作为解法介入"
+        "→ 对应卖点与真实使用过程 → 可见的改善 → 一个行动引导。开头先给可拍到的具体问题，"
+        "商品不能在第一句话直接硬推；一个痛点只对应一个卖点。禁止把普通不便夸成严重后果或制造焦虑。\n"
+    ),
+    "contrast": (
+        "【前后对比·结构合同】必须建立同一人物/场景/角度/光线下可比较的“使用前”和“使用后”，"
+        "中间必须有商品使用过程作为转折，再总结差异并 CTA。旁白只交代条件、过程和差异，"
+        "不能写成长篇原理讲解；不得跳过过程、换条件伪造差异，或同时对比多个维度。\n"
+    ),
+    "review": (
+        "【测评验证·结构合同】必须按顺序完成：提出要验证什么 → 说清测评标准/条件 → 完整展示测试动作"
+        "→ 呈现画面可观察的结果 → 给出带适用边界的结论和 CTA。开场必须是疑问或待验证目标，"
+        "不是直接夸商品；每个卖点都要改写成可演示的验证项目。禁止虚构检测数据、权威结论、用户评价或编造缺点。\n"
+    ),
+    "scene": (
+        "【场景种草·结构合同】必须按顺序完成：先交代人物、具体时间和地点 → 场景中自然出现需求"
+        "→ 商品以正在被使用的方式入场 → 写出该场景里的真实体验 → 轻度推荐/软 CTA。第一镜先建立场景，"
+        "不要一上来抛痛点提问或商品参数；商品不能像硬广道具摆拍，卖点要融入动作与感受。\n"
+    ),
+    "promo": (
+        "【促销抢购·结构合同】必须按顺序完成：真实福利/价格钩子（仅在资料明确时）→ 商品基础价值"
+        "→ 优惠适用的规格、条件与领取方式 → 真实的时间/库存信息（仅在资料明确时）→ 明确购买入口和操作 CTA。"
+        "若资料没有价格、优惠、库存或时间，就不得捏造或制造抢购氛围，改为商品价值与查看详情的引导。\n"
+    ),
+    "knowledge": (
+        "【知识分享·结构合同】必须按顺序完成：提出一个选购/使用问题或常见误区 → 给出清楚的判断方法"
+        "→ 用实物细节、步骤或正确/错误示范讲清方法 → 商品在方法建立后作为符合该标准的实例出现"
+        "→ 总结选择建议和自然 CTA。只讲 1–3 个知识点，商品不能在开头就变成卖点罗列。"
+        "禁止虚构数据、标准、专业身份或功效承诺，也不能只靠人物口头科普而没有画面演示。\n"
+    ),
+}
+
+STRUCTURE_STAGES: dict[str, tuple[str, ...]] = {
+    "pain": ("具体问题与不便", "商品介入和真实用法", "可见改善与行动引导"),
+    "contrast": ("使用前的基准状态", "商品使用过程", "使用后的同条件状态", "差异总结与行动引导"),
+    "review": ("测评目标与标准", "完整测试过程", "可观察的结果", "适用结论与行动引导"),
+    "scene": ("具体时间、地点与人物状态", "场景中的自然需求", "商品被使用着入场", "真实体验与软 CTA"),
+    "promo": ("真实福利或商品价值钩子", "商品与适用场景", "明确的优惠条件和购买方法", "入口与操作 CTA"),
+    "knowledge": ("问题或常见误区", "判断方法或核心结论", "实物步骤或正反示范", "商品承接、选择建议与 CTA"),
+}
+
 # 设定卡人物 key → 中文(与前端 WIZ_PERSONA_LABEL / 模板 coerce_persona 对齐)。
 PERSONA_LABELS: dict[str, str] = {
     "urban": "都市白领女性",
@@ -177,6 +222,68 @@ def coerce_combo(fmt: str | None, structure: str | None) -> tuple[str, str]:
     return fmt, structure
 
 
+# 与前端 script-setup.ts CATEGORY_COMBOS 对齐。改这里时两边一起改。
+_CATEGORY_COMBOS: dict[str, dict[str, str]] = {
+    "美妆个护": {"format": "oral", "structure": "contrast", "persona": "bestie"},
+    "食品饮料": {"format": "oral", "structure": "scene", "persona": "bestie"},
+    "服饰鞋包": {"format": "oral", "structure": "scene", "persona": "urban"},
+    "家居日用": {"format": "oral", "structure": "pain", "persona": "mom"},
+    "数码家电": {"format": "oral", "structure": "review", "persona": "reviewer"},
+    "母婴玩具": {"format": "oral", "structure": "pain", "persona": "mom"},
+    "运动健康": {"format": "oral", "structure": "scene", "persona": "urban"},
+    "珠宝配饰": {"format": "oral", "structure": "scene", "persona": "ceo"},
+    "汽车用品": {"format": "oral", "structure": "review", "persona": "reviewer"},
+    "宠物用品": {"format": "oral", "structure": "pain", "persona": "mom"},
+    "图书文教": {"format": "oral", "structure": "knowledge", "persona": "urban"},
+    "五金农资": {"format": "oral", "structure": "review", "persona": "reviewer"},
+    "其他商品": {"format": "oral", "structure": "pain", "persona": "urban"},
+}
+_CATEGORY_KEYWORD_RULES: tuple[tuple[tuple[str, ...], dict[str, str]], ...] = (
+    (("美妆", "护肤", "彩妆", "面膜", "精华", "洗护", "个护"), _CATEGORY_COMBOS["美妆个护"]),
+    (("保健", "营养", "膳食", "益生菌", "维生素"), {"format": "oral", "structure": "review", "persona": "reviewer"}),
+    (("数码", "3c", "电子", "电器", "手机", "耳机", "相机", "工具", "家电"), _CATEGORY_COMBOS["数码家电"]),
+    (("食品", "零食", "饮料", "咖啡", "茶", "生鲜", "酒"), _CATEGORY_COMBOS["食品饮料"]),
+    (("服饰", "女装", "男装", "服装", "鞋", "包", "配饰", "内衣"), _CATEGORY_COMBOS["服饰鞋包"]),
+    (("家居", "家纺", "收纳", "厨具", "清洁", "日用"), _CATEGORY_COMBOS["家居日用"]),
+    (("母婴", "宝宝", "儿童", "玩具"), _CATEGORY_COMBOS["母婴玩具"]),
+    (("宠物",), _CATEGORY_COMBOS["宠物用品"]),
+    (("户外", "露营", "运动", "健身"), _CATEGORY_COMBOS["运动健康"]),
+)
+_FALLBACK_COMBO: dict[str, str] = {"format": "oral", "structure": "pain", "persona": "urban"}
+
+
+def recommend_script_setup(category: str = "", title: str = "") -> dict:
+    """按商品品类推荐表现形式 × 视频结构 × 人物 × 时长。用户随时可改。"""
+    category = (category or "").strip()
+    title = (title or "").strip()
+    combo = _FALLBACK_COMBO
+    reason = "商品信息不足,先给一组最通用的"
+    if category in _CATEGORY_COMBOS:
+        combo = _CATEGORY_COMBOS[category]
+        reason = f"按「{category}」品类推荐"
+    else:
+        haystack = f"{category} {title}".lower()
+        for keywords, item in _CATEGORY_KEYWORD_RULES:
+            hit = next((word for word in keywords if word.lower() in haystack), None)
+            if hit:
+                combo = item
+                reason = f"按「{hit}」品类推荐"
+                break
+    # 自动推荐只给口播；Vlog / 短剧留给用户自己选。
+    fmt, structure = coerce_combo("oral", combo["structure"])
+    persona = coerce_persona(combo.get("persona") or "urban")
+    if persona not in PERSONA_LABELS:
+        persona = "urban"
+    duration = max(FORMAT_DEFAULT_DURATION[fmt], STRUCTURE_MIN_DURATION[structure])
+    return {
+        "format": fmt,
+        "structure": structure,
+        "persona": persona,
+        "duration": duration,
+        "reason": reason,
+    }
+
+
 def narration_limit(duration: int) -> int:
     """这一镜旁白的字数上限:秒数 × 5.2,且不超过硬上限 78。"""
     return max(1, min(NARRATION_CHARS_HARD_CAP, int(duration * NARRATION_CHARS_PER_SECOND)))
@@ -367,8 +474,12 @@ _CREATIVE_DIRECTION = """
 
 ## 出片感与转化感（硬性创作方向）
 
+**先服从用户选定的视频结构。** 用户选的是痛点解决、前后对比、测评验证、场景种草、促销抢购或知识分享中的哪一种，
+就只走该结构合同的顺序与说服方式；不能因为“更好写”而擅自换成痛点开场、测评口吻或场景氛围。
+下列通用写法仅用于增强已选结构，和结构合同冲突时一律以结构合同为准。
+
 你的脚本不能像商品详情页、说明书或主播念稿；它必须让用户在第一秒看到一个
-**正在发生的具体瞬间**。先选一个最能代表目标人群的「时间 + 地点 + 小麻烦/小欲望」，
+**正在发生的具体瞬间**。先选一个最能代表典型使用者的「时间 + 地点 + 小麻烦/小欲望」，
 全片只围绕这一个情境推进。不要在一条短视频里罗列所有卖点。
 
 ### 每一镜都必须有推进，禁止平铺直叙
@@ -487,8 +598,7 @@ def _product_context(project, selling_point_ids: list[str] | None, persona: str
         f"品牌:{product.brand or '未填写'}\n"
         f"{type_line}"
         f"品类:{product.category or '未填写'}\n"
-        f"目标人群:{product.target_audience or '未填写'}\n"
-        f"人物设定:{persona_text or '未指定,按目标人群自行定一个具体身份'}\n"
+        f"人物设定:{persona_text or '未指定,按商品信息自行定一个具体身份'}\n"
         f"人物口吻:{persona_brief or '按人物设定自己定一个具体身份,不要用万能主播腔'}\n"
         f"商品描述:{desc}\n"
         f"规格:{specs_block}\n"
@@ -572,6 +682,7 @@ def build_agent_messages(
 
     suggested = plan_segment_durations(total, fmt)
     shot_n = len(suggested)
+    stage_plan = structure_stage_plan(structure, shot_n)
     extra = (user_prompt or "").strip()
     if target_index is not None and base_draft:
         existing_n = len(base_draft.get("segments") or [])
@@ -610,24 +721,10 @@ def build_agent_messages(
             f"【写法硬约束】有人声的镜口播/对白 {speech_floor}–{speech_cap} 字,允许个别镜纯画面;"
             f"visual 至少 {VISUAL_CHARS_MIN} 字。\n"
         )
-    structure_line = ""
-    if structure == "scene":
-        structure_line = (
-            "【场景种草骨架】按镜走完:场景建立(商品先不出现)→商品被使用着入场→场景因它更好/软CTA。"
-            "每镜写出具体时刻+地点+光/声音/物件。口播时每镜仍要说满字数,留白靠停顿不是靠少写字。\n"
-        )
-    elif structure == "promo":
-        structure_line = (
-            "【促销抢购合规】只可使用商品资料中明确给出的价格、优惠、库存、赠品和活动时间；"
-            "没有明确数据时，绝不写原价、最低价、限时、限量、领券或赠品。"
-            "结尾只给一个清楚的购买入口和操作动作，不制造虚假的紧迫感。\n"
-        )
-    elif structure == "knowledge":
-        structure_line = (
-            "【知识分享骨架】先提出一个与商品相关的选购/使用误区或判断方法，再用实物与步骤证明，"
-            "最后把商品作为符合该判断的具体实例。只能讲商品资料可支持的事实，"
-            "禁止编造专业身份、数据、标准或功效承诺。\n"
-        )
+    structure_line = STRUCTURE_CONTRACTS[structure]
+    stage_line = "【镜序验收】" + "；".join(
+        f"第 {index + 1} 镜：{stage}" for index, stage in enumerate(stage_plan)
+    ) + "。镜数不够时只可合并相邻步骤，顺序不得颠倒，也不得插入其他结构的主线。\n"
     combo_line = (
         f"严格按已加载的「{PRESENTATION_FORMATS[fmt]} × {VIDEO_STRUCTURES[structure]}」套路写,"
         f"不要串成别的结构的套话。\n"
@@ -646,6 +743,7 @@ def build_agent_messages(
         f"{writing_line}"
         f"{beats_line}"
         f"{structure_line}"
+        f"{stage_line}"
         f"{combo_line}"
         f"{authenticity_line}"
         f"【商品信息】\n{_product_context(project, selling_point_ids, persona)}"
@@ -685,7 +783,7 @@ def build_agent_messages(
                 "写 visual 时把拆解稿里的景别/机位/运镜/人物动作/表情/音效/背景音乐/字幕/备注"
                 "折进【声音】和【画面内容】的秒级分镜,不要压成一句画面摘要。"
                 "原稿写「无 / 不可见 / 听不清」的栏不要编造。"
-                "钩子、痛点、卖点必须能对上这份商品。"
+                "钩子、已选结构的核心步骤和商品事实必须能对上。"
                 "每镜旁白要能撑满指定时长;画面必须按秒拆分镜,够导演在 15 秒里切 3–5 刀。\n"
                 f"{head}"
                 f"{extra_block}\n"
@@ -695,7 +793,7 @@ def build_agent_messages(
             user = (
                 "【任务】全自动(模式①):仅凭上面的商品事实与前置条件,按指定的表现形式与视频结构套路"
                 "自动定镜/选 tone/造 entity/填结构骨架。"
-                "不要另起一个空主题;钩子、痛点、卖点必须能对上这份商品,而不是品类套话。"
+                "不要另起一个空主题;钩子、已选结构的核心步骤和商品事实必须能对上这份商品,而不是品类套话。"
                 "每镜旁白要能撑满 15 秒;画面必须按秒拆分镜,够导演在 15 秒里切 3–5 刀。\n"
                 f"{head}"
                 f"{extra_block}\n"
@@ -793,6 +891,20 @@ def plan_segment_durations(total_duration: int, presentation_format: str | None
     return [SEGMENT_DURATION_MAX] * count
 
 
+def structure_stage_plan(video_structure: str, segment_count: int) -> list[str]:
+    """把选中结构的完整说服路径按当前镜数连续压缩，绝不换成通用痛点模板。"""
+    stages = list(STRUCTURE_STAGES[video_structure])
+    count = max(1, int(segment_count or 1))
+    if count >= len(stages):
+        return stages + [stages[-1]] * (count - len(stages))
+    # 前面镜先承接前置步骤，最后一镜必须保留收束/CTA；每个阶段只会向相邻镜合并。
+    groups: list[list[str]] = [[] for _ in range(count)]
+    for index, stage in enumerate(stages):
+        target = min(count - 1, index * count // len(stages))
+        groups[target].append(stage)
+    return [" → ".join(group) for group in groups]
+
+
 def plan_roles(count: int) -> list[str]:
     """镜数 → role 序列。通用规则:首钩子、次痛点、末 CTA,中间全是卖点。"""
     if count <= 1:

core/backend/apps/ai/services.py

@@ -894,6 +894,20 @@ class VideoSubmissionStateUnknown(ProviderOutcomeUnknownError):
     """视频提交可能已到达供应商但未拿到可靠任务 ID；禁止自动重提。"""
 
 
+NO_EMBEDDED_CAPTIONS_REQUIREMENT = (
+    "【成片硬性规则】画面不得出现字幕、台词字幕、花字、标题、价格贴片、UI 文字或其他后期文字叠加；"
+    "只保留商品包装上参考图中原有的真实物理文字与标识。旁白和环境音可以保留，但不要把声音转成画面文字。"
+)
+
+
+def enforce_no_embedded_captions(prompt: str) -> str:
+    """所有视频入口最终汇入这里，避免某个入口漏传“无字幕”导致模型自行加花字。"""
+    base = (prompt or "").strip()
+    if NO_EMBEDDED_CAPTIONS_REQUIREMENT in base:
+        return base
+    return f"{base}\n{NO_EMBEDDED_CAPTIONS_REQUIREMENT}".strip()
+
+
 def execute_routed_video_submit(
     *,
     task: AITask,
@@ -913,6 +927,7 @@ def execute_routed_video_submit(
     """异步视频只路由“提交”阶段；拿到 Provider 任务 ID 后固定该实际模型轮询。"""
     from apps.billing.pricing import quote_video_estimate
 
+    prompt = enforce_no_embedded_captions(prompt)
     references = list(reference_images or [])
     routed_content_items = list(content_items) if content_items is not None else None
     pricing_refs = list(pricing_references or [])
@@ -2693,7 +2708,7 @@ def build_storyboard_frame_prompt(project, segment, extra_prompt: str = "") -> s
         补充=(("\n" + extra_prompt.strip()) if extra_prompt else ""),
     )
     cleaned = "\n".join(line for line in rendered.split("\n") if line.strip())
-    return _apply_storyboard_output_ratio(cleaned, project)
+    return enforce_no_embedded_captions(_apply_storyboard_output_ratio(cleaned, project))
 
 
 def build_video_segment_prompt(project, video_segment, scene, refs, user_prompt: str = "") -> str:
@@ -2729,7 +2744,7 @@ def build_video_segment_prompt(project, video_segment, scene, refs, user_prompt:
         脚本=(script_text or f"第 {video_segment.sort_order + 1} 段"),
         时长=video_segment.target_duration_seconds,
     )
-    return "\n".join(line for line in rendered.split("\n") if line.strip())
+    return enforce_no_embedded_captions("\n".join(line for line in rendered.split("\n") if line.strip()))
 
 
 def ensure_storyboard_shots(project) -> list:
@@ -2904,7 +2919,7 @@ def build_storyboard_frame_prompt_refs(project, segment, refs: list[dict], extra
         补充=(("\n" + extra_prompt.strip()) if extra_prompt else ""),
     )
     cleaned = "\n".join(line for line in rendered.split("\n") if line.strip())
-    return _apply_storyboard_output_ratio(cleaned, project)
+    return enforce_no_embedded_captions(_apply_storyboard_output_ratio(cleaned, project))
 
 
 def _is_transient_error(exc: Exception) -> bool:

core/backend/apps/ai/test_script_duration_combo.py

@@ -13,6 +13,7 @@ from apps.ai.script_agent import (
     DEFAULT_TOTAL_DURATION,
     PRESENTATION_FORMATS,
     SEGMENT_DURATION_MAX,
+    STRUCTURE_CONTRACTS,
     TOTAL_DURATION_MAX,
     TOTAL_DURATION_MIN,
     VIDEO_STRUCTURES,
@@ -35,6 +36,8 @@ from apps.ai.script_agent import (
     parse_visual_beats,
     plan_roles,
     plan_segment_durations,
+    recommend_script_setup,
+    structure_stage_plan,
 )
 
 ALL_TOTALS = [15, 30, 45, 60]
@@ -81,6 +84,16 @@ class SegmentPlanningTests(SimpleTestCase):
             self.assertEqual(plan[0], "钩子")
             self.assertEqual(plan[-1], "CTA")
 
+    def test_structure_stage_plan_preserves_selected_structure_order_when_compressed(self):
+        self.assertEqual(
+            structure_stage_plan("knowledge", 2),
+            ["问题或常见误区 → 判断方法或核心结论", "实物步骤或正反示范 → 商品承接、选择建议与 CTA"],
+        )
+        self.assertEqual(
+            structure_stage_plan("contrast", 4),
+            ["使用前的基准状态", "商品使用过程", "使用后的同条件状态", "差异总结与行动引导"],
+        )
+
 
 class SegmentDurationFittingTests(SimpleTestCase):
     def test_uneven_durations_are_forced_to_15(self):
@@ -265,7 +278,6 @@ class PromptAssemblyTests(SimpleTestCase):
         product.brand = "AirBrand"
         product.business_type = "ecommerce"
         product.category = "美妆"
-        product.target_audience = "油皮"
         product.description = "夜间修护"
         product.specs = {"price": "99", "source": "product-create-upload"}
         product.selling_points.all.return_value = _QS([sp])
@@ -290,6 +302,7 @@ class PromptAssemblyTests(SimpleTestCase):
         self.assertNotIn("一句话主题扩写", user)
         self.assertIn("【补充要求】节奏紧凑,适合短视频投放", user)
         self.assertIn("控油12小时", user)
+        self.assertNotIn("目标人群:", user)
         self.assertIn("都市白领女性", user)
         self.assertIn("价格:99", user)
         self.assertIn("每镜必须 15 秒", user)
@@ -301,6 +314,38 @@ class PromptAssemblyTests(SimpleTestCase):
         self.assertIn("禁止使用「唯一缺点", user)
         self.assertIn("测试过程里的意外发现", user)
 
+    def test_selected_video_structure_injects_its_own_contract(self):
+        expected_titles = {
+            "pain": "【痛点解决·结构合同】",
+            "contrast": "【前后对比·结构合同】",
+            "review": "【测评验证·结构合同】",
+            "scene": "【场景种草·结构合同】",
+            "promo": "【促销抢购·结构合同】",
+            "knowledge": "【知识分享·结构合同】",
+        }
+        for structure, expected_title in expected_titles.items():
+            with self.subTest(structure=structure):
+                messages = build_agent_messages(
+                    project=self._project(),
+                    mode="auto",
+                    user_prompt="",
+                    selling_point_ids=["1"],
+                    base_draft=None,
+                    aspect_ratio="9:16",
+                    total_duration=30,
+                    presentation_format="oral",
+                    video_structure=structure,
+                    persona="urban",
+                )
+                user = messages[1]["content"]
+                self.assertIn(STRUCTURE_CONTRACTS[structure], user)
+                self.assertIn(expected_title, user)
+                self.assertIn("【镜序验收】", user)
+                self.assertIn(structure_stage_plan(structure, 2)[0], user)
+                for other, other_title in expected_titles.items():
+                    if other != structure:
+                        self.assertNotIn(other_title, user)
+
     def test_video_digest_prompt_maps_camera_and_sound(self):
         digest = (
             "片名：《雨夜归家》\n视频类型：剧情短片\n"
@@ -338,7 +383,7 @@ class PromptAssemblyTests(SimpleTestCase):
             persona="urban",
         )
         user = messages[1]["content"]
-        self.assertIn("场景种草骨架", user)
+        self.assertIn("【场景种草·结构合同】", user)
         self.assertIn("口播每镜必须说满", user)
         self.assertIn("一共 3 镜", user)
         self.assertIn("秒级分镜", user)
@@ -550,3 +595,45 @@ class SkillLoadingTests(SimpleTestCase):
         self.assertIn("每镜固定 15 秒", text)
         self.assertIn("商品原词", text)
         self.assertNotIn("单镜 4–15 秒，允许不等长", text)
+
+
+class RecommendScriptSetupTests(SimpleTestCase):
+    def test_exact_category_beauty(self):
+        rec = recommend_script_setup("美妆个护", "精华")
+        self.assertEqual(rec["format"], "oral")
+        self.assertEqual(rec["structure"], "contrast")
+        self.assertEqual(rec["persona"], "bestie")
+        self.assertIn("美妆个护", rec["reason"])
+
+    def test_digital_uses_reviewer(self):
+        rec = recommend_script_setup("数码家电")
+        self.assertEqual(rec["format"], "oral")
+        self.assertEqual(rec["structure"], "review")
+        self.assertEqual(rec["persona"], "reviewer")
+
+    def test_food_uses_oral_scene_bestie(self):
+        rec = recommend_script_setup("食品饮料", "轻醒咖啡")
+        self.assertEqual(rec["format"], "oral")
+        self.assertEqual(rec["structure"], "scene")
+        self.assertEqual(rec["persona"], "bestie")
+
+    def test_legacy_keyword_fallback(self):
+        rec = recommend_script_setup("服饰内衣", "连衣裙")
+        self.assertEqual(rec["format"], "oral")
+        self.assertEqual(rec["structure"], "scene")
+        self.assertEqual(rec["persona"], "urban")
+
+    def test_never_recommends_vlog_or_drama(self):
+        for category in (
+            "美妆个护", "食品饮料", "服饰鞋包", "家居日用", "数码家电",
+            "母婴玩具", "运动健康", "珠宝配饰", "汽车用品", "宠物用品",
+            "图书文教", "五金农资", "其他商品",
+        ):
+            rec = recommend_script_setup(category)
+            self.assertEqual(rec["format"], "oral", category)
+
+    def test_empty_falls_back_to_generic(self):
+        rec = recommend_script_setup("", "")
+        self.assertEqual(rec["format"], "oral")
+        self.assertEqual(rec["structure"], "pain")
+        self.assertEqual(rec["persona"], "urban")

core/backend/apps/ai/test_video_caption_policy.py

@@ -0,0 +1,16 @@
+from django.test import SimpleTestCase
+
+from apps.ai.services import NO_EMBEDDED_CAPTIONS_REQUIREMENT, enforce_no_embedded_captions
+
+
+class VideoCaptionPolicyTests(SimpleTestCase):
+    def test_always_appends_no_embedded_captions_requirement(self):
+        prompt = enforce_no_embedded_captions("一位用户在厨房展示商品")
+
+        self.assertIn(NO_EMBEDDED_CAPTIONS_REQUIREMENT, prompt)
+        self.assertIn("一位用户在厨房展示商品", prompt)
+
+    def test_requirement_is_not_duplicated(self):
+        prompt = enforce_no_embedded_captions(NO_EMBEDDED_CAPTIONS_REQUIREMENT)
+
+        self.assertEqual(prompt.count(NO_EMBEDDED_CAPTIONS_REQUIREMENT), 1)

core/backend/apps/ai/test_video_segment_routing.py

@@ -10,7 +10,7 @@ from django.utils import timezone
 
 from apps.accounts.models import Team, TeamMember, User
 from apps.ai.models import AITask, ModelConfig, ModelProvider
-from apps.ai.services import poll_video_segment, submit_video_segment
+from apps.ai.services import NO_EMBEDDED_CAPTIONS_REQUIREMENT, poll_video_segment, submit_video_segment
 from apps.assets.models import Asset
 from apps.billing.models import CreditAccount, CreditLedger
 from apps.products.models import Product
@@ -140,6 +140,8 @@ class VideoSegmentRoutingTests(TestCase):
         self.assertEqual(self.ledger_count(task, CreditLedger.Type.RESERVE), 1)
         self.assertEqual(self.ledger_count(task, CreditLedger.Type.CHARGE), 0)
         self.assertEqual(self.ledger_count(task, CreditLedger.Type.RELEASE), 0)
+        submitted_prompt = provider.create_video_task.call_args.kwargs["prompt"]
+        self.assertIn(NO_EMBEDDED_CAPTIONS_REQUIREMENT, submitted_prompt)
 
     def test_second_submit_does_not_create_another_task(self):
         primary = self.model(self.provider("video-primary-once", 20), "video-primary-once", outbound=True, default=True)

core/backend/apps/ai/tryon_prompt.py

@@ -488,7 +488,7 @@ _KEYWORDS: tuple[tuple[ProductKind, tuple[str, ...]], ...] = (
 _GENERIC_WEARABLE_HINTS = (
     "围巾", "披肩", "领带", "领结", "袜", "腰带", "皮带", "手套", "护膝", "护腕", "服饰", "服装", "穿戴", "衣帽", "鞋服",
 )
-_BROAD_WEARABLE_CATEGORIES = ("服饰内衣", "服饰", "服装", "女装", "男装", "配饰")
+_BROAD_WEARABLE_CATEGORIES = ("服饰内衣", "服饰鞋包", "服饰", "服装", "女装", "男装", "配饰")
 _USER_RELATION_RULES: tuple[tuple[ProductKind, tuple[str, ...]], ...] = (
     (ProductKind.HANDHELD, ("手持", "拿着", "拿在手", "握住")),
     (ProductKind.WEARABLE_GENERIC, ("穿上", "穿着", "换上", "佩戴", "戴上", "戴着")),

core/backend/apps/products/models.py

@@ -12,6 +12,11 @@ class Product(TeamOwnedModel):
         ECOMMERCE = "ecommerce", "电商"
         LOCAL_LIFE = "local_life", "本地生活"
 
+    ECOMMERCE_CATEGORIES = (
+        "美妆个护", "食品饮料", "服饰鞋包", "家居日用", "数码家电", "母婴玩具", "运动健康",
+        "珠宝配饰", "汽车用品", "宠物用品", "图书文教", "五金农资", "其他商品",
+    )
+
     title = models.CharField(max_length=255)
     brand = models.CharField(max_length=128, blank=True)
     # 类目(电商/本地生活)≠品类(美妆/食品)。类目驱动节点屏蔽与后续 skill 挂载。

core/backend/apps/projects/serializers.py

@@ -423,11 +423,12 @@ class QuickCreateJobSerializer(serializers.ModelSerializer):
                 if version is not None and is_playable_video(version.asset):
                     video_url = _asset_preview_url(version.asset)
                     break
-        first_shot = next(
-            (shot for shot in project.storyboard_shots.all() if shot.adopted_version_id),
-            None,
-        )
-        poster_url = _asset_preview_url(first_shot.adopted_version.asset) if first_shot else ""
+        product = getattr(project, "product", None)
+        cover = getattr(product, "cover_asset", None) if product is not None else None
+        poster_url = _asset_preview_url(cover) if cover else ""
+        if not poster_url and product is not None:
+            first_image = next(iter(product.images.all()), None)
+            poster_url = _asset_preview_url(first_image.asset) if first_image else ""
         script = next((version for version in project.script_versions.all() if version.is_adopted), None)
         script_meta = (script.metadata or {}) if script is not None else {}
         format_label = {"oral": "口播展示", "drama": "短剧演绎", "vlog": "Vlog种草"}.get(

core/backend/apps/projects/services/quick_create.py

@@ -15,7 +15,13 @@ from django.db import connections, transaction
 from django.utils import timezone
 
 from apps.ai.models import AITask, ModelConfig
-from apps.ai.script_agent import stream_script_agent
+from apps.ai.script_agent import (
+    PRESENTATION_FORMATS,
+    VIDEO_STRUCTURES,
+    persona_label,
+    recommend_script_setup,
+    stream_script_agent,
+)
 from apps.ai.services import (
     collect_video_review_blockers,
     create_export_job,
@@ -364,6 +370,17 @@ def _consume_script_agent(job: QuickCreateJob) -> None:
     if model_config is None:
         raise ValueError(f"{QUICK_SCRIPT_MODEL_NAME} is not configured")
 
+    wizard = dict((project.metadata or {}).get("wizard") or {})
+    product = project.product
+    rec = recommend_script_setup(
+        getattr(product, "category", "") or "",
+        getattr(product, "title", "") or "",
+    )
+    fmt = wizard.get("presentation_format") or rec["format"]
+    structure = wizard.get("video_structure") or rec["structure"]
+    persona = wizard.get("persona") or rec["persona"]
+    format_label = PRESENTATION_FORMATS.get(fmt, fmt)
+    structure_label = VIDEO_STRUCTURES.get(structure, structure)
     error_detail = ""
     stream = stream_script_agent(
         project=project,
@@ -371,15 +388,16 @@ def _consume_script_agent(job: QuickCreateJob) -> None:
         model_config=model_config,
         mode="auto",
         user_prompt=(
-            "请根据商品名称与商品参考图信息，自动推荐最适合的"
-            f"{settings['total_duration']}秒、{settings['aspect_ratio']}画幅带货方案。"
+            "请根据商品名称、品类与商品参考图信息，按推荐的"
+            f"{format_label} × {structure_label} × {persona_label(persona)}，"
+            f"生成{settings['total_duration']}秒、{settings['aspect_ratio']}画幅带货方案。"
         ),
         aspect_ratio=settings["aspect_ratio"],
         total_duration=settings["total_duration"],
-        presentation_format="oral",
-        video_structure="pain",
+        presentation_format=fmt,
+        video_structure=structure,
         entry_source="ai",
-        persona="reviewer",
+        persona=persona,
     )
     # 豆包思考模型会连续吐几百帧 reasoning。原先每帧 refresh_from_db，远程 MySQL
     # 一抖动就会把生成器关掉，模型调用被误记成 stream aborted (client disconnected)。

core/backend/apps/projects/test_quick_create.py

@@ -76,6 +76,7 @@ class QuickCreateApiTests(TestCase):
                 "aspect_ratio": "9:16",
                 "resolution": "720p",
                 "total_duration": "30",
+                "category": "食品饮料",
             },
             format="multipart",
         )
@@ -83,10 +84,14 @@ class QuickCreateApiTests(TestCase):
         self.assertEqual(response.status_code, 202)
         product = Product.objects.get(title="轻醒咖啡")
         self.assertEqual(product.images.count(), 2)
+        self.assertEqual(product.category, "食品饮料")
         self.assertEqual(product.cover_asset_id, product.images.order_by("sort_order").first().asset_id)
         project = Project.objects.get(product=product)
         self.assertTrue(project.metadata["quick_create"])
         self.assertEqual(project.metadata["wizard"]["total_duration"], 30)
+        self.assertEqual(project.metadata["wizard"]["presentation_format"], "oral")
+        self.assertEqual(project.metadata["wizard"]["video_structure"], "scene")
+        self.assertEqual(project.metadata["wizard"]["persona"], "bestie")
         self.assertEqual(project.metadata["wizard"]["resolution"], "720p")
         self.assertEqual(project.metadata["wizard"]["video_model_config_id"], str(video_model_id))
         self.assertEqual(project.stages.count(), 5)
@@ -106,6 +111,30 @@ class QuickCreateApiTests(TestCase):
             {str(image.asset_id) for image in product.images.order_by("sort_order")},
         )
 
+    @patch("apps.projects.services.quick_create.get_quick_script_model", return_value=object())
+    @patch("apps.projects.views.get_default_model")
+    @patch("apps.projects.tasks.advance_quick_create_task.apply_async")
+    @patch("apps.projects.views.require_worker_task")
+    @patch("apps.projects.views._store_uploaded_asset")
+    def test_submit_requires_product_category(self, store_asset, require_worker_task, enqueue, get_model, get_quick_model):
+        store_asset.side_effect = self._uploaded_asset
+        get_model.side_effect = [
+            object(),
+            SimpleNamespace(id=uuid.uuid4(), name="seedance", display_name="Seedance", metadata={}),
+        ]
+        response = self.client.post(
+            "/api/projects/quick-create/",
+            {
+                "name": "无品类咖啡",
+                "images": [SimpleUploadedFile("front.png", b"png", content_type="image/png")],
+            },
+            format="multipart",
+        )
+        self.assertEqual(response.status_code, 400)
+        self.assertIn("品类", str(response.data["detail"]))
+        self.assertFalse(Product.objects.filter(title="无品类咖啡").exists())
+        enqueue.assert_not_called()
+
     @patch("apps.projects.services.quick_create.get_quick_script_model", return_value=object())
     @patch("apps.projects.views.get_default_model", return_value=object())
     @patch("apps.projects.tasks.advance_quick_create_task.apply_async")
@@ -129,6 +158,7 @@ class QuickCreateApiTests(TestCase):
             {
                 "name": "复用商品图",
                 "source_product_id": str(source.id),
+                "category": "美妆个护",
                 "aspect_ratio": "9:16",
                 "resolution": "720p",
                 "total_duration": "15",
@@ -140,6 +170,7 @@ class QuickCreateApiTests(TestCase):
         store_asset.assert_not_called()
         product = Product.objects.get(title="复用商品图")
         self.assertEqual(product.images.count(), 1)
+        self.assertEqual(product.category, "美妆个护")
         self.assertEqual(product.images.first().asset_id, source_asset.id)
         self.assertEqual(response.data["product_images"][0]["asset_id"], str(source_asset.id))
 
@@ -651,6 +682,52 @@ class QuickCreateCoordinatorTests(TestCase):
         self.assertEqual(self.job.status, QuickCreateJob.Status.RUNNING)
         enqueue.assert_called_once()
 
+    @patch("apps.projects.services.quick_create.get_quick_script_model")
+    @patch("apps.projects.services.quick_create.stream_script_agent")
+    def test_consume_script_uses_category_recommendation(self, stream_fn, get_model):
+        get_model.return_value = self.model
+        self.product.category = "食品饮料"
+        self.product.title = "轻醒咖啡"
+        self.product.save(update_fields=["category", "title", "updated_at"])
+        self.project.metadata = {"wizard": {"total_duration": 30, "aspect_ratio": "9:16"}}
+        self.project.save(update_fields=["metadata", "updated_at"])
+        stream_fn.return_value = iter([])
+        with self.assertRaises(ValueError):
+            _consume_script_agent(self.job)
+        kwargs = stream_fn.call_args.kwargs
+        self.assertEqual(kwargs["presentation_format"], "oral")
+        self.assertEqual(kwargs["video_structure"], "scene")
+        self.assertEqual(kwargs["persona"], "bestie")
+        self.assertEqual(kwargs["total_duration"], 30)
+        self.assertIn("口播", kwargs["user_prompt"])
+        self.assertIn("场景种草", kwargs["user_prompt"])
+        self.assertIn("闺蜜种草", kwargs["user_prompt"])
+
+    @patch("apps.projects.services.quick_create.get_quick_script_model")
+    @patch("apps.projects.services.quick_create.stream_script_agent")
+    def test_consume_script_keeps_wizard_combo(self, stream_fn, get_model):
+        get_model.return_value = self.model
+        self.product.category = "食品饮料"
+        self.product.save(update_fields=["category", "updated_at"])
+        self.project.metadata = {
+            "wizard": {
+                "presentation_format": "oral",
+                "video_structure": "pain",
+                "persona": "reviewer",
+                "total_duration": 15,
+                "aspect_ratio": "9:16",
+            }
+        }
+        self.project.save(update_fields=["metadata", "updated_at"])
+        stream_fn.return_value = iter([])
+        with self.assertRaises(ValueError):
+            _consume_script_agent(self.job)
+        kwargs = stream_fn.call_args.kwargs
+        self.assertEqual(kwargs["presentation_format"], "oral")
+        self.assertEqual(kwargs["video_structure"], "pain")
+        self.assertEqual(kwargs["persona"], "reviewer")
+        self.assertEqual(kwargs["total_duration"], 15)
+
     @patch("apps.projects.services.quick_create.get_quick_script_model")
     @patch("apps.projects.services.quick_create.stream_script_agent")
     def test_consume_script_does_not_hit_db_on_every_sse_frame(self, stream_fn, get_model):

core/backend/apps/projects/views.py

@@ -21,6 +21,7 @@ from apps.ai.script_agent import (
     SEGMENT_DURATION_MIN,
     combo_keys,
     coerce_total_duration,
+    recommend_script_setup,
     stream_script_agent,
 )
 from apps.ai.services import (
@@ -607,6 +608,11 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
                 return Response({"detail": "请至少上传一张商品图片"}, status=status.HTTP_400_BAD_REQUEST)
         if len(reused_assets) + len(uploads) > 9:
             return Response({"detail": "商品图片最多上传9张"}, status=status.HTTP_400_BAD_REQUEST)
+        category = str(request.data.get("category") or "").strip()
+        if not category:
+            return Response({"detail": "请选择商品品类"}, status=status.HTTP_400_BAD_REQUEST)
+        if category not in Product.ECOMMERCE_CATEGORIES:
+            return Response({"detail": "请选择有效的商品品类"}, status=status.HTTP_400_BAD_REQUEST)
 
         with transaction.atomic():
             assets = list(reused_assets)
@@ -628,6 +634,7 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
                 created_by=request.user,
                 title=name,
                 business_type=Product.BusinessType.ECOMMERCE,
+                category=category,
                 cover_asset=assets[0],
             )
             ProductImage.objects.bulk_create(
@@ -641,6 +648,7 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
                     for index, asset in enumerate(assets)
                 ]
             )
+            rec = recommend_script_setup(category, name)
             project = Project.objects.create(
                 team=team,
                 created_by=request.user,
@@ -657,9 +665,9 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
                         "video_model_config_id": str(video_model.id),
                         "video_model_name": video_model.name,
                         "video_model_label": video_model.display_name,
-                        "presentation_format": "oral",
-                        "video_structure": "pain",
-                        "persona": "reviewer",
+                        "presentation_format": rec["format"],
+                        "video_structure": rec["structure"],
+                        "persona": rec["persona"],
                     },
                 },
             )
@@ -697,9 +705,10 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
 
     def _quick_job_queryset(self):
         return (
-            QuickCreateJob.objects.select_related("project__product", "project__timeline")
+            QuickCreateJob.objects.select_related("project__product", "project__product__cover_asset", "project__timeline")
             .prefetch_related(
                 "project__product__images__asset__files",
+                "project__product__cover_asset__files",
                 "project__script_versions",
                 "project__base_asset_groups",
                 "project__storyboard_shots__adopted_version__asset__files",