11) Recipe 系统：把“100+ 生成器”工程化成可维护的工作流库

新增入口只新增 recipe 配置；核心引擎与执行管线保持稳定，可 A/B、可运营、可回放。

11.1 Recipe 配置示例

{
  "recipe_id": "r_url_to_shorts_v1",
  "name": "URL → 9:16 Shorts",
  "inputs": {
    "url": "string",
    "language": "zh-CN",
    "target_length_sec": 45,
    "tone": "energetic"
  },
  "defaults": {
    "aspect": "9:16",
    "fps": 30,
    "caption_style": "bold_pop",
    "cut_density_per_min": 22,
    "bgm_target_lufs": -18,
    "vo_target_lufs": -14
  },
  "dag": [
    {"id":"fetch", "type":"fetch_url"},
    {"id":"extract", "type":"extract_keypoints", "deps":["fetch"]},
    {"id":"script", "type":"write_script", "deps":["extract"]},
    {"id":"scenes", "type":"scene_plan", "deps":["script"]},
    {"id":"stock", "type":"stock_search", "deps":["scenes"]},
    {"id":"tts", "type":"tts_generate", "deps":["script"]},
    {"id":"caps", "type":"caption_generate_align", "deps":["tts"]},
    {"id":"mix", "type":"mix_audio", "deps":["tts"]},
    {"id":"preview", "type":"render_preview", "deps":["stock","caps","mix"]},
    {"id":"final", "type":"render_final", "deps":["preview"]}
  ],
  "quality_gates": [
    {"rule":"no_black_frames", "threshold": 0},
    {"rule":"caption_safe_area_violation", "threshold": 0},
    {"rule":"audio_clipping_ratio", "threshold": 0.001}
  ]
}

回到顶部

12) 模型路由与降级：让系统在峰值、失败与成本压力下仍可交付

12.1 降级阶梯（Fallback Ladder）

VideoGen 失败 / 排队爆炸时：
1) 复用已有素材（stock b-roll）替代生成镜头
2) 降分辨率（preview 先 540p）
3) 缩短片段时长（先出核心段）
4) 使用更快模型（质量稍降）
5) 仅输出静帧 + 动效（Ken Burns / zoom）过渡

回到顶部

13) 协作与多人编辑：从一开始就按“可合并的工程文件”设计

13.2 Patch 合并策略（简单但好用）

Merge rules (example):
- 字幕样式类字段：可自动合并（last-write-wins）
- 时间线结构类操作（move/split）：若目标 clip 已被移动 → 冲突
- 语义编辑（rewrite scene script）：若 scene 文本未变 → 自动合并；否则提示用户选择版本

回到顶部

14) 存储与缓存：决定“实时感”和单位成本的关键

14.3 Render Segment 键设计（示例）

segment_key = sha256(
  clip_ids + timing + effects +
  asset_hashes + style_hash +
  resolution + fps + codec
)

回到顶部

15) 音频链路：旁白、BGM、ducking、响度标准化（决定“专业感”）

15.3 Ducking 伪代码（简化）

// given vo_segments: [{start,end}, ...]
for each frame t:
  if t in vo_segments:
    bgm_gain = lerp(bgm_gain, target=-12dB, attack=80ms)
  else:
    bgm_gain = lerp(bgm_gain, target=0dB, release=220ms)

回到顶部

16) 字幕系统：生成、对齐、排版、安全区（短视频爆款必备）

16.3 Safe Area 规则示例（9:16）

safe_top = 0.10 * height
safe_bottom = 0.18 * height
caption_box.y must be in [safe_top, height - safe_bottom - box.h]

回到顶部

17) API / BFF 设计：支持 WebSocket 进度回推与可恢复任务

17.2 WebSocket 事件示例

{
  "type": "job_progress",
  "project_id": "p_123",
  "job_id": "job_88",
  "progress": 0.62,
  "stage": "render_preview",
  "artifacts": [{"type":"video_preview","uri":".../prev.mp4"}],
  "revision": 43
}

回到顶部

18) 测试与评估：让“生成系统”可回归、可持续迭代

18.3 Timeline Diff 示例（概念）

diff:
- clip_7 start: 11.2s → 10.6s
- clip_7 speed: 1.0 → 1.15
- aud_vo_2 asset: as_tts_21 → as_tts_33
- cap_2 text changed (len 18 → 12)

回到顶部

19) 安全与合规：上传素材、版权与隔离（商业化绕不过）

回到顶部

20) 基础设施与成本：把单位成本压到可持续（规模化必修课）

20.3 “先交付、后精炼”的产能模型

Phase 1: draft (fast, cheap)  → 用户确认方向
Phase 2: refine (targeted)     → 只对不满意段落做重操作
Phase 3: final export (heavy)  → 只在用户点击导出时触发

回到顶部

附录（续）：更完整的对象模型与事件溯源（Event Sourcing）

事件溯源建议结构

{
  "event_id": "ev_10092",
  "project_id": "p_123",
  "rev": 43,
  "type": "PATCH_APPLIED",
  "actor": {"kind":"assistant","id":"sys"},
  "ops": ["op_9f31","op_aa02"],
  "job_id": "job_88",
  "ts": "2026-01-05T10:12:11Z",
  "summary": "加快 clip_7，替换旁白并重对齐字幕"
}

常见事件类型

回到顶部

21) LLM 规划器（Planner）：从自然语言到可执行 Patch 的“可控链路”

关键不是让模型更聪明，而是让它在可控轨道上输出：先定位（Locate）再规划（Plan）再校验（Validate）再执行（Execute）。

21.3 多阶段规划（更稳）

Stage A: Normalize
- 规范化结构：{target_range?, target_scene?, change_type, constraints}

Stage B: Locate (tool + rules + LLM)
- 先用索引/规则定位候选，再让 LLM 选择最可能目标
- 输出：target_ids + confidence + fallback question(if needed)

Stage C: Plan
- 生成 patch（ops），每个 op 显式绑定对象 ID

Stage D: Validate & Repair
- schema 校验失败 → 回灌错误让模型修复（最多 N 次）
- 语义校验失败（越界/引用不存在）→ 重定位或请求用户澄清

Stage E: Execute
- 进入 DAG + Queue

Stage F: Summarize
- 同步：改动说明、预览何时可见、最终导出是否额外时间

回到顶部

22) 定位系统（Locator）：把“那段/第二点/12 秒”映射到工程对象

现象级体验的隐藏关键：用户永远模糊表达，但系统必须精确命中对象。

22.1 三种定位信号（建议都实现）

22.2 Locator 工具接口（示例）

{
  "type": "locate_targets",
  "input": {
    "project_id":"p_123",
    "query":"把第二个观点那段节奏加快一点",
    "hints":{"selected_ids":[], "cursor_time": 11.8}
  },
  "output": {
    "candidates":[
      {"type":"scene","id":"s2","range":[6.5,18.0],"score":0.86},
      {"type":"clip","id":"clip_7","range":[11.2,14.3],"score":0.63}
    ],
    "best":{"type":"scene","id":"s2"},
    "confidence":0.86
  }
}

回到顶部

23) 风格系统（Style System）：把“好看”变成模板、约束与可复用资产

23.3 Style Profile 示例

{
  "style_id": "style_bold_pop_v2",
  "caption": {
    "font": "Inter-Bold",
    "stroke_px": 6,
    "shadow": true,
    "highlight": {"mode":"keyword", "max_per_line":2}
  },
  "motion": {
    "kenburns_strength": 0.35,
    "transition": {"type":"whip", "duration_ms":240},
    "cut_density_per_min": 24
  },
  "audio": {
    "vo_target_lufs": -14,
    "bgm_target_lufs": -18,
    "ducking_db": -12,
    "duck_attack_ms": 80,
    "duck_release_ms": 220
  },
  "safe_area": {"top": 0.10, "bottom": 0.18}
}

回到顶部

24) Prompt / Tooling：把“提示词”产品化成可审计、可回放、可回归

不要把 prompt 写死在代码里：版本化+回放才能持续迭代。

24.1 Prompt 版本化（建议）

{
  "prompt_id": "planner_patch_v6",
  "purpose": "Chat → Patch",
  "schema": "PatchSchema@1.2",
  "system": "你是视频工程规划器，只输出 JSON ...",
  "fewshots": ["..."],
  "constraints": ["不得输出自然语言", "必须绑定 clip_id", "..."],
  "created_at": "2026-01-05",
  "owner": "video-platform"
}

回到顶部

25) 内容治理与策略：把风险前置到 Planning 与导出阶段

25.3 Policy 示例（概念）

{
  "policy_id": "org_brand_policy_01",
  "allowed_stock_sources": ["sourceA","sourceB"],
  "allowed_models": ["fast_video_v1","hq_video_v2"],
  "require_license_metadata": true,
  "caption_safe_area_enforced": true,
  "banned_terms": ["..."],
  "watermark": {"enabled": false}
}

回到顶部

26) 现象级增长飞轮：工程系统如何反哺产品增长

回到顶部

27) 全链路蓝图：把系统拆成可立即开工的模块清单

回到顶部

附录（再续）：失败模式与防护（Failure Modes）清单

回到顶部

28) 索引与检索：让“找素材/找句子/找片段”像数据库一样快

对话式编辑效率来自“快速定位目标”。依赖：文本索引、时间索引、语义索引三件套。

28.3 Transcript 索引结构示例

{
  "project_id": "p_123",
  "utterances": [
    {"seg_id":"cap_2","text":"因此我们可以…","start":11.2,"end":14.0},
    {"seg_id":"cap_3","text":"接下来讲原因…","start":14.0,"end":18.0}
  ],
  "inverted": {
    "因此": ["cap_2"],
    "原因": ["cap_3"]
  }
}

回到顶部

29) B-roll 选择引擎：比“生成镜头”更省钱、更稳的爆款路径

大量爆款短视频镜头来自 stock + 动效 + 字幕，而不是全程视频生成。

29.3 选择策略：先稳后花

For each scene:
1) try stock_topK (cheap, fast)
2) if no good match → use still image + motion (Ken Burns)
3) if still weak AND user wants "更电影感" → trigger video generation for that scene only

回到顶部

30) 节奏引擎（Pacing Engine）：把“好看”变成算法与约束

“爽感”来自节奏：切镜密度、字幕更新频率、动效强度、鼓点对齐。

30.3 Pacing 调整 patch 示例

{
  "ops": [
    {"op_id":"op_p1","type":"set_pacing_profile","target":{"scene_id":"s1"},
     "params":{"avg_shot_len":1.2,"cut_density_per_min":30}},
    {"op_id":"op_p2","type":"set_pacing_profile","target":{"scene_id":"s2"},
     "params":{"avg_shot_len":2.4,"cut_density_per_min":18}}
  ]
}

回到顶部

31) 多比例导出（9:16 / 16:9 / 1:1）：不是“裁剪”而是“重排版”

多比例如果只是 crop，会导致主体跑出画面、字幕遮挡。正确做法：维护布局层（layout layer）。

31.3 Layout patch 示例

{
  "ops": [
    {"op_id":"op_l1","type":"set_clip_framing",
     "target":{"clip_id":"clip_7"},
     "params":{
       "aspect":"9:16",
       "anchor":"center",
       "scale":1.25,
       "pan":{"x":0.08,"y":-0.02}
     }}
  ]
}

回到顶部

32) 编辑器实现：Timeline UI 如何与 Patch/Job 系统强一致

轨道编辑器是 Project/DSL 的可视化前端：所有编辑都变成 patch，统一走同一条链路。

32.3 提交 patch 的典型流程

onUserDragClip(clip_id, newStart):
  patch = {base_revision, ops:[{type:"move_clip", target:{clip_id}, params:{start:newStart}}]}
  applyOptimistic(patch)
  POST /projects/:id/patch
  WS receive revision:
    if ok → commit optimistic
    if conflict → rollback + show diff

回到顶部

33) 性能与体验：端到端要做到“像实时”，工程上要做到“可渐进”

33.3 渐进式交付（Progressive Delivery）

t=0s   : show plan summary + estimated cost
t=1-3s : preview of changed segment ready
t=3-10s: captions/alignment refined
later  : final export ready (notify)

回到顶部

34) 对话式编辑的关键 UX 回路：把不确定性变成“可选择”

34.3 A/B 方案输出结构（示例）

{
  "options": [
    {"label":"A-加速","patch":{...}, "effect":"speed 1.15x, keep content"},
    {"label":"B-剪短","patch":{...}, "effect":"remove filler, shorten 2.0s"}
  ],
  "recommend": "A",
  "reason": "保持信息完整，节奏更快"
}

回到顶部

附录（续 4）：可直接开工的最小 Worker 列表与接口契约

Worker 清单（MVP）

统一 Step 接口契约（建议）

input:
{
  "job_id":"job_88",
  "step_id":"step_render_preview_1",
  "project_id":"p_123",
  "revision": 43,
  "params": {...},
  "artifacts_in": [{"artifact_id":"..."}],
  "policy_id":"org_policy_01"
}

output:
{
  "status":"SUCCEEDED",
  "artifacts_out":[{"artifact_id":"as_prev_91","type":"video","uri":"..."}],
  "metrics":{"latency_ms": 2400},
  "usage_report":{"credits": 3.2, "cache_hit": true}
}

回到顶部

35) 供应商与模型抽象（Provider Abstraction）：可换、可混用、可审计

35.3 Provider Registry 示例

{
  "providers": [
    {
      "id":"video_fast_v1",
      "type":"videogen",
      "caps":{"max_sec":6,"max_res":"720p","ref_image":false,"fps":[24,30]},
      "sla":{"p50_ms":7000,"p95_ms":18000},
      "cost":{"credits_per_sec_720p":2.4}
    },
    {
      "id":"video_hq_v2",
      "type":"videogen",
      "caps":{"max_sec":10,"max_res":"1080p","ref_image":true,"fps":[24,30]},
      "sla":{"p50_ms":25000,"p95_ms":60000},
      "cost":{"credits_per_sec_1080p":6.8}
    }
  ]
}

回到顶部

36) 一致性（Consistency）：角色一致、风格一致、时间一致的抓手

一致性=版本化 + 稳定 ID + 约束 + 缓存 + 可回放。

36.3 一致性指纹（用于缓存/回放/审计）

{
  "fingerprint": {
    "recipe":"r_url_to_shorts@2",
    "style":"style_bold_pop@2",
    "character":"char_7@5",
    "providers":{"videogen":"video_hq_v2","tts":"tts_vendor_x"},
    "seed_policy":"stable_seed",
    "policy":"org_policy_01"
  }
}

回到顶部

37) 渲染实现细节：从 Project 到 Render Graph / Segment Pipeline

37.1 Render Graph（概念）

Nodes:
- Decode(asset)
- Transform(scale/crop/pan/blur)
- Composite(layers)
- CaptionRender(text, style, safe_area)
- AudioMix(vo, bgm, ducking, loudness)
- Encode(h264/av1, bitrate)

Edges:
- timeline order
- overlay layers
- audio sidechain for ducking

回到顶部

38) 取消与恢复：让用户敢点“生成”，也敢点“取消”

38.3 可恢复关键：Artifact Catalog

{
  "job_id":"job_88",
  "artifacts": [
    {"id":"as_tts_33","type":"audio","from_step":"tts","status":"ready"},
    {"id":"as_caps_10","type":"captions","from_step":"caps","status":"ready"},
    {"id":"as_prev_91","type":"video_preview","from_step":"preview","status":"ready"}
  ]
}

回到顶部

39) 通知系统：Final 后台跑，但用户不迷路

WS 实时更新 + 离线通知（站内/邮件/Push）闭环。

39.3 通知 payload 示例

{
  "event":"final_ready",
  "project_id":"p_123",
  "job_id":"job_88",
  "revision": 43,
  "assets":[{"type":"final_video","uri":".../final.mp4"}],
  "cta":"Open Export"
}

回到顶部

40) 企业与 API：把能力“服务化”，同时保持隔离、审计与 SLA

40.3 Webhook 示例

{
  "webhook_event":"job_succeeded",
  "job_id":"job_88",
  "project_id":"p_123",
  "revision": 43,
  "artifacts":[{"type":"final_video","uri":".../final.mp4"}],
  "usage_report":{"credits_committed": 42.6}
}

回到顶部

41) 运营面板（Ops Dashboard）：把质量、成本、队列、转化统一到一个地方

41.1 核心视图

回到顶部

附录（续 5）：端到端时序图（文字版）

User: "把第二段节奏加快，并把旁白更简洁"
  ↓
Client:
  - send chat message + cursor_time/selected_ids
  ↓
BFF:
  - fetch project digest + snapshot(range) + policy + budget
  - call locator → candidates(scene s2)
  - call planner → patch(base_rev=43, ops=[...])
  - run validator → ok
  - run estimator → reserve credits
  - create job(job_90) + enqueue
  ↓
Orchestrator:
  - build DAG( tts → caps → preview )
  - dispatch steps with priority(tier, preview first)
  ↓
Workers:
  - tts_generate → as_tts_40 + timing + usage
  - caption_align → as_caps_12 + usage
  - render_preview(DirtyRange) → as_prev_102 + usage
  ↓
BFF:
  - apply patch to project(rev=44)
  - WS push: preview_ready + revision=44 + preview uri
  ↓
User:
  - sees preview, asks "再快一点"
  - loop again
  ↓
User clicks Export:
  - enqueue render_final
  - notify when final_ready

回到顶部

42) 数据契约与 Schema：让系统可扩展、可回归、可协作

42.3 PatchSchema（片段）示意

{
  "PatchSchema": "1.2",
  "project_id": "string",
  "base_revision": "int",
  "ops": [
    {
      "op_id": "string",
      "type": "enum",
      "target": "object",
      "params": "object",
      "cost_hint": "object?"
    }
  ]
}

回到顶部

43) 质量门槛（Quality Gates）工程实现：把“翻车点”变成自动阻断

导出前硬检查：失败则不给 final 或要求修复，最好带自动修复 patch。

43.3 Gate 示例：字幕安全区检查

{
  "gate":"caption_safe_area",
  "status":"FAIL",
  "evidence":[
    {"seg_id":"cap_12","time":[8.2,9.6],"box":{"x":0.12,"y":0.86,"w":0.78,"h":0.10},
     "violation":"bottom_out_of_safe_area"}
  ],
  "suggestion":"将字幕上移 8% 或缩小字号",
  "auto_repair_patch":{
    "ops":[{"op_id":"op_fix_1","type":"adjust_caption_layout",
            "target":{"track_id":"t_c1"},
            "params":{"y_offset":-0.08,"scale":0.95}}]
  }
}

回到顶部

44) 自动剪辑策略（Auto-editing）：把“默认就好看”固化成规则库

44.1 默认剪辑规则（Playbook）

44.2 Playbook 工程形式

{
  "playbook_id":"pb_shorts_default_v3",
  "rules":[
    {"id":"r1","when":"silence_gt_ms:300","do":["remove_silence","re_align_captions"]},
    {"id":"r2","when":"hook_density_low","do":["reorder_hook","increase_cut_density"]},
    {"id":"r3","when":"cta_scene","do":["caption_emphasis","motion_boost"]}
  ]
}

回到顶部

45) 分镜生成（Storyboarding）：Scene → Shot 中间层决定可编辑性

分镜层让系统能回答：讲什么、有哪些镜头、镜头对应哪句旁白，从而支持局部重算与语义编辑。

45.3 Storyboard 示例

{
  "scene_id":"s2",
  "range":[6.5,18.0],
  "summary":"解释原因",
  "shots":[
    {"shot_id":"sh_21","dur":2.0,"asset_pref":"stock","keywords":["原因","分析"],"motion":"zoom_in"},
    {"shot_id":"sh_22","dur":2.4,"asset_pref":"stock","keywords":["对比","数据"],"motion":"pan_left"},
    {"shot_id":"sh_23","dur":1.8,"asset_pref":"still+motion","keywords":["总结"],"motion":"kenburns"}
  ],
  "bindings":{
    "vo_seg_ids":["vo_5","vo_6"],
    "cap_seg_ids":["cap_9","cap_10","cap_11"]
  }
}

回到顶部

46) 可解释性（Explainability）：让用户知道“你到底改了什么”

46.3 Explain payload 示例

{
  "revision": 44,
  "target": {"scene_id":"s2","range":[6.5,18.0]},
  "changes":[
    {"type":"speed_change","clip_id":"clip_7","from":1.0,"to":1.15},
    {"type":"remove_silence","range":[12.8,13.4]},
    {"type":"rewrite_vo","audio_id":"aud_vo_2","delta_chars":-42}
  ],
  "preview_dirty_range":[11.2,14.3]
}

回到顶部

47) 爆款 Playbooks：把“从输入到爆款”的默认策略沉淀成组合包

回到顶部

附录（续 6）：自动修复（Auto-repair）策略库示例

质量门槛失败后，尽量自动修复而不是让用户猜怎么改；修复 patch 也要进入版本历史并可撤销。

{
  "auto_repairs":[
    {
      "when":"caption_safe_area_fail",
      "fix":[
        {"type":"adjust_caption_layout","params":{"y_offset":-0.08}},
        {"type":"scale_caption","params":{"scale":0.95}},
        {"type":"reduce_caption_chars","params":{"max_chars_per_line":12}}
      ]
    },
    {
      "when":"audio_clipping_fail",
      "fix":[
        {"type":"normalize_loudness","params":{"target_lufs":-14}},
        {"type":"limit_true_peak","params":{"dbtp":-1.0}}
      ]
    },
    {
      "when":"black_frames_detected",
      "fix":[
        {"type":"replace_asset","params":{"strategy":"fallback_stock"}},
        {"type":"add_background_fill","params":{"mode":"blur_fill"}}
      ]
    }
  ]
}

回到顶部

48) Op Catalog：完整的“视频编辑原子操作”集合

Op Catalog 是 DSL 的指令集。建议每个 op 都有：参数 schema、成本估算、undo、dirty range 计算。

48.1 时间线与结构类

[
  "insert_clip", "remove_clip", "move_clip", "duplicate_clip",
  "trim_clip", "split_clip", "merge_clips",
  "set_clip_speed", "set_clip_volume", "set_clip_opacity",
  "set_transition", "set_track_order", "set_layer_zindex",
  "set_scene_range", "create_scene", "delete_scene"
]

48.2 字幕与文本类

[
  "generate_captions", "align_captions", "rewrite_captions",
  "set_caption_style", "set_caption_safe_area", "adjust_caption_layout",
  "highlight_keywords", "set_caption_language", "translate_captions",
  "fix_caption_overflow", "set_caption_animation"
]

48.3 音频类（VO/BGM/SFX）

[
  "tts_generate", "replace_voiceover", "rewrite_voiceover_text",
  "select_bgm", "replace_bgm", "mix_audio", "set_ducking",
  "normalize_loudness", "limit_true_peak", "add_sfx", "remove_sfx",
  "fade_in", "fade_out"
]

48.4 视觉动效与布局类

[
  "set_clip_framing", "set_pan_zoom", "set_kenburns",
  "add_overlay", "remove_overlay", "set_overlay_position",
  "add_blur_fill", "set_global_filter", "set_color_grade",
  "set_motion_profile", "set_pacing_profile"
]

48.5 生成与资产类（高成本）

[
  "stock_search", "replace_asset",
  "generate_image_asset", "generate_video_asset",
  "regenerate_shot", "regenerate_scene",
  "create_character_profile", "update_character_profile",
  "lock_character", "lock_style"
]

回到顶部

49) 语义编辑 → 时间线映射：把“改观点/改结构”变成可执行剪辑

三层模型：Script Layer → Storyboard Layer → Timeline Layer，通过 bindings 将语义改动安全落地到局部 patch。

49.3 示例：让第二观点更清楚

1) rewrite script for scene s2 (shorter, clearer)
2) tts_generate new VO for s2
3) regenerate captions for s2
4) adjust pacing: slightly slower for s2
5) keep visuals mostly, but add 1 stock shot illustrating key term

Patch ops (high-level):
- rewrite_scene_script(s2)
- replace_voiceover(s2)
- align_captions(s2)
- set_pacing_profile(s2, avg_shot_len=2.6)
- insert_clip(after shot sh_22)

回到顶部

50) Toolchain：从“聊天”到“剪辑”的可扩展工具链

LLM 不应承担检索/计算/索引查询等硬事实工作；应编排工具并输出 patch。

50.1 推荐工具清单（可做内部 SDK）

50.2 Tool-first Planner 伪代码

// Planner.run(utterance, project_id)
targets = locate_targets(utterance, hints)
snap = get_project_snapshot(project_id, targets.range)
policy = get_policy(project_id)
cost = estimate_cost(utterance, snap, policy)

patch = llm_generate_patch(utterance, snap, policy, cost)
validate_patch(patch) || patch = llm_repair_patch(errors)

dirty = compute_dirty_range(patch, snap)
job_id = enqueue_job(patch, dirty, priority=tier)

return {patch, job_id, dirty, cost, summary}

回到顶部

51) A/B 实验框架：让 Recipe / Style / Prompt 迭代有数据闭环

现象级产品迭代不是拍脑袋：用实验比较成功率、导出率、成本。

51.3 分流结构示例

{
  "experiment_id":"exp_recipe_shorts_v2",
  "variants":[
    {"name":"A","recipe":"r_url_to_shorts@2","style":"bold_pop@2"},
    {"name":"B","recipe":"r_url_to_shorts@3","style":"bold_pop@2"}
  ],
  "allocation":{"A":0.5,"B":0.5},
  "unit":"user_id",
  "start":"2026-01-05",
  "metrics":["first_draft_success","export_rate","credits_per_min"]
}

回到顶部

52) 个性化（Personalization）：用默认偏好提升成功率

52.3 Preference Profile 示例

{
  "user_id":"u_01",
  "defaults":{
    "style":"style_bold_pop@2",
    "voice":"voice_zh_female_03",
    "aspect":"9:16",
    "playbook":"pb_shorts_default_v3",
    "pacing":{"cut_density_per_min":24}
  }
}

回到顶部

附录（续 7）：Undo（撤销/重做）策略

撤销/重做若只做前端，会在异步任务/协作/多端下崩溃。建议从 revision rollback 起步。

Undo 两种实现

Inverse Patch 示例（概念）

{
  "op":"move_clip",
  "target":{"clip_id":"clip_7"},
  "params":{"start":10.6},
  "inverse":{
    "op":"move_clip",
    "target":{"clip_id":"clip_7"},
    "params":{"start":11.2}
  }
}

回到顶部