sources_v5.json
55.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
{
"task": "T-mtd7r60r",
"agent": "innovator-hunt (Hunter)",
"round": "retrain v5 HUNT (两篇飞书官方文档首次入库 + 两个 GitHub 社区 skill)",
"crawl_date": "2026-08-28",
"fetched_at_utc": "2026-08-28T17:05:00Z - 2026-08-28T17:29:00Z",
"time_precision_note": "crawl_time 为分钟级精度(同一批次共用一个时刻),非逐请求秒级记录",
"method": "飞书走 /app/bin/lark-cli docs +fetch(bot 身份,先 outline 后 markdown,全文一次取全未分段);GitHub 走 api.github.com git/trees 拿全量路径 + raw.githubusercontent.com 逐文件抓取;sha256 本地计算;行号为缓存落盘文件内 1-based 行号",
"grading_legend": {
"A": "官方一手(MiniMax 官方飞书文档),可逐字回查 cache_v5/",
"B": "社区一手实践(GitHub 公开仓全文已落盘),工程规范可参考,与官方冲突处不采信",
"C": "个人方法论 / 无官方出处的自创词汇,只登记不采信",
"R": "复用自 v4 已固化信源,本轮禁复抓,逐字出处走 train/hunt/sources_v4.json"
},
"sources": [
{
"id": "s01",
"grade": "A",
"title": "飞书官方 · 🐻MiniMax H3 模型 - 使用手册(本轮最高优先级,此前从未进过任何一轮语料)",
"url": "https://vrfi1sk8a0.feishu.cn/wiki/FIWjwgL33ipnkekzk30crmKUnIh",
"revision_id": 1297,
"document_id": "ZRFZd89oMoGTyrx2JXWcpPq1n7f",
"bytes": 147058,
"sha256": "48a7a693c70b4e8fd6567a2f4d3ce049ec3966577a3b2ee0e4db556c752b98b6",
"crawl_time": "2026-08-28T17:06:00Z",
"fetch_via": "lark-cli docs +fetch --doc-format markdown(bot 身份,ok:true)",
"cache": "cache_v5/feishu_doc1_H3使用手册_FIWjwgL33ipnkekzk30crmKUnIh.md",
"structure": "一、模型定位 / 二、模型功能信息 / 三、模型亮点能力&示例(1.1-1.9 + 2.1-2.3 + 3.1-3.4)/ 四、如何更好地使用 H3(1. 整体公式·四要素·镜头怎么拆·三类模式·容易踩的坑;2. 特定风格·提示词示意;3. 最佳方式)",
"quotes": [
{
"line": 35,
"text": "<td><b>输出时长</b></td><td>4-15秒</td> … <td><b>输出帧率</b></td><td>24 FPS</td> … <td><b>输出声音</b></td><td>所有结果全部带声音输出,原生双声道</td> … <td><b>提示词字数上限</b></td><td>不超过7000字符</td>",
"note": "模型硬参数表(A 级一手):4-15秒、24FPS、原生双声道、提示词 <=7000 字符;另有 全能参考入口 图片<=9张 / 视频<=3段(单段[2,15]s, 总<=15s) / 音频<=3段且必须配图或视频 / 混合输入总上限 12 个文件;首尾帧入口 图片 0/1/2 张;768p 已开放、1440p 官方推荐版"
},
{
"line": 71,
"text": "官方小Tips:如果你在处理文字时遇到乱码问题,可尝试将文字以图片形式提供给 H3 作为参考图,并在提示词中明确告诉 H3——\"这张内容必须按\"图片\"来理解,不能按\"文字\"来处理。\" 通常能有效降低乱码情况。",
"note": "★官方一手新增条款,v1-v4 全部没有:文字乱码 → 转参考图 + 显式声明按图理解",
"verbatim": false,
"verbatim_lines": "71",
"reformat_note": "原文是 HTML table 内嵌引号嵌套,text 为保义转写;逐字以 cache :71 为准"
},
{
"line": 97,
"text": "[0 秒-2 秒] 高角度俯拍。… [2 秒-4 秒] 平滑变焦至她的右臂。… [4 秒-7 秒] 摄像机平滑地绕到她的左侧。… [7 秒-8.5 秒] 摄像机拉回至中景。… [8.5 秒-10 秒] 底部出现加载 (LOADING) 条。… [10 秒-15 秒] 当她起身时,完整的环境在她周围加载。",
"note": "★v4 master 案例(六时间窗 0-2/2-4/4-7/7-8.5/8.5-10/10-15)的官方一手同型确认。v4 只能靠 GitHub 镜像 co-op 模板(s13)间接证明,本轮升为 A 级一手逐字"
},
{
"line": 186,
"text": "开源用户Prompt指南:<cite doc-id=\"Vztzwstuqi3MQKkxs8QcGmxinxe\" file-type=\"wiki\" title=\"MiniMax H3 - 视频扩写 Prompt 写作全能指南\" type=\"doc\"></cite>",
"note": "官方自己把 doc2 定位为 doc1 的 Prompt 指南出向引用,两篇文档互为正源"
},
{
"line": 195,
"text": "完整提示词 = 参考素材说明 + 核心创意 + 画面过程说明",
"note": "★官方整体公式(三段式)"
},
{
"line": 200,
"text": "画面过程说明 按时间轴或故事线分段写,每个分镜/时间段包含两部分:想要(画面景别+内容+运镜+动作+台词+音效)/ 不想要(视频里不要出现什么)… 非叙事性音乐:N/A 或者使用英文格式:\"non_diegetic_music\": N/A … 0-3 秒:全景,女子(@图片1)从画面左侧缓步走入樱花庭院(@图片2)… 3-8 秒:切镜到女子的中景… 8-12 秒:切镜到特写,剑光一闪,慢动作… 非叙事性音乐:N/A",
"note": "四要素表逐字出处;官方同时给出中文字段名(非叙事性音乐)与英文字段名(non_diegetic_music)两种写法 → 直接印证 v4 的习语分轨必要性;且给出了第三种时间分段形态「N-M 秒行式」",
"verbatim": false,
"verbatim_lines": "200",
"reformat_note": "doc1:200 为单行超大 HTML table,text 是其中多格的紧凑摘录;逐字以 cache :200 为准"
},
{
"line": 200,
"text": "写清即可;注意环绕运镜,建议用 truck left+pan right 或者 truck right+pan left,而非直接说环绕运镜",
"note": "★官方一手新增:环绕运镜的替代组合写法(与 doc2:123 运镜表含 Arc Shot 构成官方内部张力,见 knowhow_v5 §冲突 C-7)"
},
{
"line": 200,
"text": "尽量少写比喻句,多写看得见的画面 … 因为 H3 更擅长理解明确的画面指令,而不是抽象比喻",
"note": "官方一手:禁比喻/意会句条款"
},
{
"line": 206,
"text": "台词的长短和画面变化内容需要合乎每个shot的长短逻辑,尽量不要出现一个3s shot说很大一段话的情况,会非常影响效果",
"note": "★台词-镜头时长对齐条款;官方原注「很多口型问题会来自于这个」"
},
{
"line": 208,
"text": "H3 能够响应跨shot (J-cut,L-cut)的台词响应,只要明确写出一句台词跨了哪些shot就可以响应!",
"note": "官方一手:跨镜台词官方术语 J-cut / L-cut"
},
{
"line": 209,
"text": "对于一个shot而言如果是画面内说话人角色,需写清楚是哪个角色说的,如果是画面外说话人,写清楚是画外说话人。如果发生了画内外因为切镜造成的转化的情况可以明确写出来",
"note": "画外/画内说话人身份显式条款;:211-213 给出「画外音 Wake up Wake up → 切镜到中年妇女近景,她是画外音的主人,她继续说道 It's time to go to school!」完整官方示例"
},
{
"line": 216,
"text": "切镜的时候可明确写出切镜到什么景别,具体主体是之前的哪个角色,有助于跨镜头一致性的保持",
"note": "切镜一致性条款(跨镜身份保持的官方写法)"
},
{
"line": 222,
"text": "上传多模态素材融合(多模态参考)… 每个素材都要写清它的角色:@图片1 → 人物参考(锁脸)/ @视频1 → 动作参考(锁动作)/ @音频1 → 节奏/情绪参考 … 上传图片生成视频(图生视频)只上传 1 张图:说清楚是首帧还是尾帧;上传 2 张图(首+尾帧):H3 不会自动加切镜,只补两帧之间的动作、光影、声音 … 纯文字生成视频:多用「大全景交代空间 + 中景承载动作 + 特写强调细节」的分层写法",
"note": "三类模式官方写法;2 张首尾帧图不会自动加切镜是一条易错点",
"verbatim": false,
"verbatim_lines": "222",
"reformat_note": "doc1:222 为单行 HTML table,text 是三格紧凑摘录;逐字以 cache :222 为准"
},
{
"line": 231,
"text": "想一镜到底但写了很多分镜 → 全文保持一段情节描述,删掉【镜头 N】结构",
"note": "官方坑表:一镜到底与分镜结构互斥 → 分镜结构的取舍由「是否切镜」决定,不是默认加",
"verbatim": false,
"verbatim_lines": "231",
"reformat_note": "原文为 markdown 表格行「| 想一镜到底但写了很多分镜 | 全文保持… |」,text 用 → 连接;逐字以 cache :231 为准"
},
{
"line": 237,
"text": "【画面过程描述】0〜3 秒- 正向:… - 反向:… 3〜6 秒- 正向:… 6〜10 秒- 正向:… 10〜13 秒- 正向:… 13〜15 秒- 正向:…",
"note": "官方「特定风格·提示词示意」完整 15s 案例:第四种时间分段形态「0〜3 秒」全角波浪线 + 每段「正向/反向」两栏(正向=要什么,反向=不要什么)"
}
],
"used_for": [
"模型硬参数升 A 级",
"整体公式/四要素官方正源",
"v4 六时间窗案例官方一手确认",
"乱码转图条款(新增)",
"时间分段多制式盘点(见 knowhow_v5 §冲突 C-1)"
]
},
{
"id": "s02",
"grade": "A",
"title": "飞书官方 · MiniMax H3 - 视频扩写 Prompt 写作全能指南(T2VA/I2VA/FL2VA/L2VA + 全能参考六段式 + 标签体系)",
"url": "https://vrfi1sk8a0.feishu.cn/wiki/Vztzwstuqi3MQKkxs8QcGmxinxe",
"revision_id": 252,
"document_id": "DhOudwGrAoykcRxjBOoc4pLonTf",
"bytes": 41607,
"sha256": "caa721f6140cc06f1450c961d38372059191fc5e106cef787249a8098f8c0e06",
"crawl_time": "2026-08-28T17:09:00Z",
"fetch_via": "lark-cli docs +fetch --doc-format markdown(bot 身份,ok:true)",
"cache": "cache_v5/feishu_doc2_Prompt写作全能指南_Vztzwstuqi3MQKkxs8QcGmxinxe.md",
"structure": "引言 / 一、文生视频·关键帧生视频(1 任务介绍;2 最终 Prompt 结构;3 关键帧写进多模态描述;4 共用三主体段落;5 Cases 1-4)/ 二、全能参考(1 整体结构;2 引用标签与定义;3 summary;4 retention_analysis;5 detailed_description;6 overall_soundscape 与 non_diegetic_music;7 完整示例)",
"quotes": [
{
"line": 28,
"text": "- **T2VA**:根据文字构建完整的视听时间线。\n- **I2VA**:T2VA 主体 + 首帧 instruction + 从首帧继续发展的画面路径。\n- **FL2VA**:T2VA 主体 + 首尾帧 instruction + 从首帧到尾帧的连续路径。\n- **L2VA**:T2VA 主体 + 尾帧 instruction + 从合理的前置状态收敛到尾帧的路径。",
"note": "四类 VA 路径的官方定义"
},
{
"line": 42,
"text": "For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.",
"note": "I2VA 固定 instruction 逐字模板"
},
{
"line": 48,
"text": "How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot N) aligns with the S.SS-second mark of the target video.",
"note": "FL2VA 固定 instruction 逐字模板"
},
{
"line": 54,
"text": "How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the S.SS-second mark of the target video.",
"note": "L2VA 固定 instruction 逐字模板"
},
{
"line": 57,
"text": "其中,`N` 是最后一个实际镜头的编号;`S.SS` 是视频有效时长,固定保留两位小数。instruction 必须位于最终 Prompt 的第一行,后面空一行再写主体字段。",
"note": "★instruction 位置硬规则(第一行 + 空行)+ S.SS 两位小数定义"
},
{
"line": 62,
"text": "integrated_multimodal_description: [Shot 1] ...\n\noverall_soundscape: ...\n\nnon_diegetic_music: ...",
"note": "T2VA/I2VA/FL2VA/L2VA 三主体字段官方顺序(:69-71 给出各字段职责)"
},
{
"line": 79,
"text": "推荐结构:**首帧锚点 → 动作起点 → 连续变化 → 结果或反应**。",
"note": "I2VA 四段推荐结构(:87 FL2VA:首帧状态→可观察的中间变化→差异逐步缩小→尾帧状态;:93 L2VA:合理前态→明确的动作与变化路径→最后一个镜头逐步收敛→尾帧落点)"
},
{
"line": 101,
"text": "`[Shot 1]` 的开头先写整体风格与初始构图。常见风格包括 `Cinematic`、`live-action`、`2D-animated`、`3D CG`、`claymation`、`watercolor` 和 `vintage film`。关键帧任务的风格应从参考图中提取;T2VA 则根据用户文字选择。",
"note": "风格词官方枚举 + 风格来源规则"
},
{
"line": 107,
"text": "第一镜不写时间戳。",
"note": "★与 v4 base-en.txt:86 的「Do not add a timestamp to the first shot」官方一手逐字对应,升级为 A 级"
},
{
"line": 111,
"text": "后续镜头使用连续编号,并在开头写递增且位于视频时长范围内的切镜时间:",
"note": "与 v4 base-en.txt:86「strictly increasing cut time」官方一手对应"
},
{
"line": 114,
"text": "[Shot 2] At 00:03.500, the camera cuts to...",
"note": "★切镜时间的官方具体格式:`At MM:SS.mmm`(三位毫秒)。v4 习语甲未固化该格式"
},
{
"line": 117,
"text": "普通切镜可使用 `the camera cuts to`、`the shot cuts to`、`the shot transitions to`、`the shot changes to` 或 `the shot switches to`,一些更加高级的切镜手法也可以灵活表述。",
"note": "切镜动词官方白名单(5 个)"
},
{
"line": 121,
"text": "一个完整的镜头运动表达由三个维度组成:**运镜类型**决定相机如何运动,**幅度**说明构图变化范围,**速度**说明变化节奏(中等幅度和常速通常省略)。",
"note": "运镜三维度官方正源(v4 引 GitHub base-en.txt:94-96,本轮升 A 级)"
},
{
"line": 123,
"text": "运镜:Zoom In / Zoom Out;Push In / Pull Out;Pan Left / Pan Right;Truck Left / Truck Right;Tilt Up / Tilt Down;Pedestal Up / Pedestal Down;Arc Shot;Tracking Shot;Static Shot;Shake Slightly / Shake Strongly;POV;Roll Clockwise / Roll Counterclockwise。幅度:with small amplitude / with large amplitude。速度:at slow speed / at fast speed",
"note": "★官方运镜词汇全表:12 组运镜 + 2 档幅度 + 2 档速度",
"verbatim": false,
"verbatim_lines": "123",
"reformat_note": "官方运镜表是 HTML table,text 把它改写为紧凑列举;逐字以 cache :123 为准"
},
{
"line": 128,
"text": "The camera pushes in with small amplitude at slow speed toward the folded letter in her hands.\nThe camera pans right with large amplitude at fast speed, revealing the open doorway.\nThe camera holds a static shot as the runner exits the frame.",
"note": "运镜写入句式的官方三例"
},
{
"line": 135,
"text": "会说话、演唱或发出画外人声的主体使用稳定编号 `(S1)`、`(S2)`、...、`(Sn)`。多个已有编号的人同时说或唱时使用组合编号比如`(S1,S2)`。同一说话人跨镜头保持同一编号;没有发声的人物不分配编号。",
"note": "★说话人编号体系官方正源(v4 线上缺陷 conv-mtclmne2 的双说话人编号问题的官方出处)"
},
{
"line": 137,
"text": "说话人第一次出现时,根据画面和声音补充稳定身份,例如人物类型、年龄、性别、是否出镜,以及音高、音色、语速或口音。说话人的称呼、编号、动作和说话方式写在 `<d>` 外;`<d>` 内只写语言标签和用户给出的实际语言内容,原词与标点逐字保留,不翻译、不改写。",
"note": "<d> 边界规则官方正源"
},
{
"line": 144,
"text": "画外音使用精确表达 `says in an off-screen voiceover`。每一段画外音 `<d>` 后立即写明画面中对应人物的嘴唇保持闭合:",
"note": "★画外音 + 闭嘴绑定条款官方正源(v4 引 base-en 相邻条款推演,本轮升 A 级逐字)"
},
{
"line": 147,
"text": "The man (S1) says in an off-screen voiceover: <d>[English] I still remember that road.</d> while his lips remain completely closed.",
"note": "画外音官方逐字示例"
},
{
"line": 150,
"text": "同一句台词或歌词跨切镜时,在两段连接位置使用 `<scenetrans>`,并明确声音跨切镜连续;视频结束前台词被截断时使用 `<cutoff>`。可使用 `continues seamlessly across the cut`、`continues uninterrupted into the next shot`、`carries over from the previous shot` 或 `remains audible across the transition` 说明跨镜头的连续性。",
"note": "<scenetrans>/<cutoff> 官方正源 + 4 个连续性表达"
},
{
"line": 154,
"text": "画面中实际可见的横幅、招牌、标识、字幕或霓虹文字使用英文双引号包围,原文与标点逐字保留,不翻译。",
"note": "画面文字条款官方正源(v4 引 base-en.txt:146,升 A 级)"
},
{
"line": 157,
"text": "A red neon sign reading \"营业中\" glows above the doorway.",
"note": "画面文字官方示例:中文招牌在英文正文里逐字保留"
},
{
"line": 162,
"text": "使用 1–4 句英文,在一个连续段落中概括贯穿整段视频的环境声、物理动作声和非语言人声 … 只有用户明确要求整段静音时才写 `N/A`。",
"note": "★overall_soundscape 官方硬约束:1–4 句英文、单段落、何时写 N/A"
},
{
"line": 170,
"text": "使用 1–3 句英文描述角色听不到、只有观众能听到的背景配乐。重点写乐器、速度、节奏和动态变化,不写抽象情绪词或配乐的情感功能。角色能听见的演唱、乐器、广播、电视或手机音乐属于场景内事件,应写入多模态描述。没有画外配乐时写 `N/A`。",
"note": "★non_diegetic_music 官方硬约束:1–3 句英文 + 禁抽象情绪词 + 画内/画外音乐分流"
},
{
"line": 242,
"text": "扩写的六个段落正文统一使用英文;只有 `<d>` 中的台词、歌词以及画面中实际可见的文字保留原始语言。",
"note": "★语言规则官方一手逐字(六段正文英文)→ 支持 v4 习语甲;与官方 handdrawn skill(s19:13 跟随用户语言)仍构成官方内部冲突,见 knowhow_v5 §冲突 C-2"
},
{
"line": 244,
"text": "`detailed_description` 应尽量详细、明确,逐镜交代当前构图、主体外观与位置、环境与光线、动作及状态变化、相机运动、当下声音,以及参考内容实际出现或生效的位置,避免只概括剧情或引用关系。",
"note": "detailed_description 详细度七要素"
},
{
"line": 250,
"text": "一份完整扩写结果由六个段落组成,按以下顺序排列:subject_definitions / summary / retention_analysis / detailed_description / overall_soundscape / non_diegetic_music",
"note": "全能参考六段官方顺序",
"verbatim": false,
"verbatim_lines": "250-260",
"reformat_note": "六段顺序原文是 markdown 表(段落/作用两列),text 折叠成单行列举;逐字以 cache :250-260 为准"
},
{
"line": 267,
"text": "| `<Subject N>` | 从参考素材中抽象出的、可在目标视频中复用或修改的可见内容 |\n| `<Picture N>` | 作为具体目标帧或镜头规划依据的参考图片 |\n| `<Video N>` | 提供剪辑源、续写起点或整段时序结构的参考视频 |\n| `<Audio N>` | 被复制或参考的音频信号 |",
"note": "四类引用标签官方定义表;:272 「某个引用标签一旦分配给一项内容 … 始终沿用同一含义」= 标签稳定性硬规则"
},
{
"line": 310,
"text": "<Picture 3> is a storyboard reference for [Shot 1] and [Shot 2], defining their viewpoint, subject placement, and shot order.",
"note": "★分镜/故事板参考图映射到镜头的官方一手逐字出处。v4 线上真实调用(conv-mtclmne2)漏认领 Picture 3 的根因条款,v4 只能引 GitHub 镜像 ref-en.txt:73,本轮升 A 级一手"
},
{
"line": 337,
"text": "`<Audio N>` 明确对应某个目标说话人时,在定义中沿用该说话人的全局编号:对应已定义主体时写作 `<Subject N> (Sx)`,否则使用稳定的声音称呼加 `(Sx)`。该编号来自目标视频统一的说话人顺序,不在音频定义中单独分配或重新编号",
"note": "<Audio N> 与 (Sx) 绑定规则"
},
{
"line": 347,
"text": "`<Video N>` 与 `<Audio N>` 各自独立编号,序号只表示它们在所属标签类别中的顺序,不表示二者之间的配对关系。因此,同一份参考视频可以同时对应 `<Video 1>` 和 `<Audio 2>`,序号不同不影响它们来自同一份素材。",
"note": "★Video/Audio 独立编号不配对(易错点)"
},
{
"line": 360,
"text": "这一段用一个简短英文段落概括目标视频及其引用关系。开头使用方括号标记任务类型:\n[reference generation] ...\n[video editing + reference generation + audio reuse] ...",
"note": "summary 开头方括号任务类型标记",
"verbatim": false,
"verbatim_lines": "360-365",
"reformat_note": "text 的第二行示例取自 :364;逐字以 cache :360-365 为准"
},
{
"line": 371,
"text": "`keyframe completion` / `reference generation` / `video editing` / `video continuation` / `audio reuse` / `audio reference`",
"note": "六种官方任务类型枚举(:372-376 逐项给使用场景)",
"verbatim": false,
"verbatim_lines": "371-376",
"reformat_note": "六任务类型原文是 markdown 表(:371-376),text 折叠成列举;逐字以 cache :371-376 为准"
},
{
"line": 378,
"text": "同一任务同时满足多个关系时,使用加号组合,且加号前后各保留一个空格,不重复同一类型。",
"note": "任务类型组合格式硬规则"
},
{
"line": 402,
"text": "| `fully_preserved` | 对应参考内容的既定作用被完整保留 | … `partially_preserved` … `attribute_transfer` … `weak_reference`",
"note": "★可见内容四关系标记(:400-405)是「格式中的固定英文取值」——该措辞逐字在同文件 :398(`<Subject N>`、`<Picture N>` 和 `<Video N>` 使用以下关系标记。这些标记是格式中的固定英文取值:)",
"verbatim_lines": "402-405"
},
{
"line": 427,
"text": "`fully_copy` | 完整源音频作为目标视频的完整最终音轨 … `partially_copy` … `reference` … `weak_reference`",
"note": "音频四关系标记"
},
{
"line": 410,
"text": "<Subject 1> (appears in [Shot 1], [Shot 3]): fully_preserved - ...\n<Picture 2> ([Shot 1] first frame): fully_preserved - ...\n<Video 1> (cut and pacing structure): weak_reference - ...",
"note": "retention_analysis 三种条目写法(:416 图片、:422 视频结构)"
},
{
"line": 455,
"text": "`[Shot 1]` 表示起始镜头,不带时间戳;后续镜头使用 `[Shot N] At MM:SS.mmm, ...` 标记切镜时间。",
"note": "全能参考 detailed_description 与 T2VA 共用同一镜头时间戳格式"
},
{
"line": 530,
"text": "| 主体字段 | `integrated_multimodal_description` | `detailed_description` |\n| 风格开头 | 写在 `[Shot 1]` 后 | 在 `[Shot 1]` 前用 1—2 句英文建立全局视觉风格 |\n| 引用信息 | 不使用全能参考模式的引用标签 | 在首次出现和实际作用位置插入 `<Subject N>`、`<Picture N>`、`<Video N>`、`<Audio N>` |",
"note": "★T2VA vs 全能参考四维差异表(第五维音频关系在 :533)",
"verbatim_lines": "530-533"
},
{
"line": 543,
"text": "生成类任务的 `detailed_description` 通常为 350—500 个英文单词。台词密集型内容以完整容纳台词时间线为先,不机械补足字数;视频编辑类内容则按源视频复杂度展开,不强制套用生成类区间。单镜头并不自动缩短篇幅,多镜头按各镜头的信息量分配细节。",
"note": "★官方字数区间:生成类 350–500 英文词 + 三条例外(台词密集/视频编辑/单镜头不缩篇幅)"
},
{
"line": 552,
"text": "the shot begins from <Picture 1>\nthe shot's keyframe corresponds to <Picture 2>\nthe shot ends on <Picture 3>",
"note": "具体帧锚点三种官方自然表达(首帧/关键帧/尾帧)",
"verbatim_lines": "552-555"
},
{
"line": 564,
"text": "<Subject 2> (S1) turns toward the woman and says, <d>[English] Last summer, I went to my grandfather's house. He talked about you.</d>",
"note": "引用标签 + 说话人编号并用的官方示例"
},
{
"line": 575,
"text": "直接复用参考音频中的台词、旁白或歌词 … `<d>` 内必须使用忠实的原始字词并保留原语言;无法辨认的部分写作 `[unclear]`,不猜测或改写。标点统一使用表达句意所需的基础书面标点,如 `,`、`.`、`?`和 `!`;删除重复波浪号、emoji、项目符号以及重复或装饰性标点。完整的陈述句、疑问句和感叹句分别在 `</d>` 前以 `.`、`?`或 `!` 结束。",
"note": "★[unclear] 规则 + 标点规范化 + 句末标点三选一(官方硬约束,v4 无)"
},
{
"line": 577,
"text": "仅参考音色、节奏、情绪或表达方式时,不应将参考音频中的原台词带入目标视频。",
"note": "audio reference 与 audio reuse 的分界"
},
{
"line": 579,
"text": "`(Sx)` 根据目标视频中的实际发声顺序统一分配。… `retention_analysis` 不写 `(Sx)`。",
"note": "(Sx) 分配范围 + retention_analysis 禁写 (Sx)"
},
{
"line": 621,
"text": "subject_definitions:\n<Subject 1> is the coffee-shop environment in <Picture 1>, featuring an exposed brick wall, an orange tufted sofa with patterned pillows, a neon sign, and a wooden coffee table.\n…\nsummary:\n[reference generation + audio reference] The target video shows <Subject 3> eating a cookie in <Subject 1>.\n…\nretention_analysis:\n<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the exposed brick wall … are retained.\n…\ndetailed_description:\nThe target video uses a realistic multi-camera sitcom style with warm indoor lighting.\n[Shot 1] A medium shot establishes <Subject 1> …\n[Shot 2] At 00:03.000, the shot cuts to a close-up of <Subject 4> (S2) …\n[Shot 3] At 00:05.000, the shot cuts to a close-up of <Subject 3> (S1) …\noverall_soundscape:\nSoft indoor coffee-shop room tone continues throughout the scene.\nnon_diegetic_music:\nN/A",
"note": "★官方完整六段示例(:620-649),可作为 v5 SKILL 的唯一权威 worked example 母本"
}
],
"used_for": [
"四类 VA 路径官方结构",
"全能参考六段 + 标签体系官方正源",
"359-430 任务类型与关系标记枚举",
"350-500 词官方字数区间",
"Picture 3 分镜映射官方一手",
"语言规则官方逐字",
"完整官方 worked example"
]
},
{
"id": "s03",
"grade": "B",
"title": "GitHub 社区仓 · chiphoton/MiniMax-H3-Codex-Drama(ComfyUI 侧 8 skill 全链路,工程化最完整)",
"url": "https://github.com/chiphoton/MiniMax-H3-Codex-Drama",
"default_branch": "main",
"crawl_time": "2026-08-28T17:18:00Z - 2026-08-28T17:25:00Z",
"tree_url": "https://api.github.com/repos/chiphoton/MiniMax-H3-Codex-Drama/git/trees/main?recursive=1",
"tree_bytes": 78321,
"tree_sha256": "55823306539962f6",
"tree_paths_total": 214,
"files_fetched": 18,
"files_total_note": "仓内其余路径为 mp4/webp/png/jpg/svg 二进制与 ComfyUI workflow json,不在本轮抓取范围",
"fetch_via": "raw.githubusercontent.com 逐文件(18/18 HTTP 200,每请求间隔 2s)",
"cache_root": "cache_v5/github/chiphoton/",
"files": [
{
"path": "README.md",
"bytes": 11880,
"sha256_16": "8b1f169968d12d86"
},
{
"path": "README_zh.md",
"bytes": 11408,
"sha256_16": "5206b941fb387c1b"
},
{
"path": "docs/turbo-vs-standard/prompt.txt",
"bytes": 812,
"sha256_16": "2c8257d2bdc0c044"
},
{
"path": "examples/exp-001-jiangnan-ancient-drama/README.md",
"bytes": 6211,
"sha256_16": "54b2cf2ea1919cc4"
},
{
"path": "skills/minimax-h3-adviser/SKILL.md",
"bytes": 7227,
"sha256_16": "b66919baf69a1665"
},
{
"path": "skills/minimax-h3-adviser/references/prompt-language.md",
"bytes": 7720,
"sha256_16": "5fe5b09be7109e9c"
},
{
"path": "skills/minimax-h3-adviser/references/workflow-map.md",
"bytes": 3340,
"sha256_16": "ec8b42f78902f181"
},
{
"path": "skills/minimax-h3-drama-producer/SKILL.md",
"bytes": 8363,
"sha256_16": "c141d4b8c4cc0d06"
},
{
"path": "skills/minimax-h3-drama-producer/references/production-workflow.md",
"bytes": 3373,
"sha256_16": "8c5dbcc08f114101"
},
{
"path": "skills/minimax-h3-drama-producer/references/visual-development.md",
"bytes": 3226,
"sha256_16": "a5138aa344b83b69"
},
{
"path": "skills/minimax-h3-drama-producer/references/profiles/base-video/shot-patterns.md",
"bytes": 428,
"sha256_16": "af2cc6baf52eb7f5"
},
{
"path": "skills/minimax-h3-drama-producer/references/profiles/base-video/storytelling.md",
"bytes": 372,
"sha256_16": "4429b873506b63bb"
},
{
"path": "skills/minimax-h3-frame-to-video/SKILL.md",
"bytes": 3017,
"sha256_16": "2caf3e535907d14a"
},
{
"path": "skills/minimax-h3-frame-to-video/references/template-and-examples.md",
"bytes": 3712,
"sha256_16": "524093504eb41e2a"
},
{
"path": "skills/minimax-h3-reference-to-video/SKILL.md",
"bytes": 3837,
"sha256_16": "334f6aef4d62cd7d"
},
{
"path": "skills/minimax-h3-reference-to-video/references/template-and-examples.md",
"bytes": 4663,
"sha256_16": "970e31b8a978462b"
},
{
"path": "skills/minimax-h3-text-to-video/SKILL.md",
"bytes": 3017,
"sha256_16": "5f42dd4bbd0a74d4"
},
{
"path": "skills/minimax-h3-text-to-video/references/template-and-examples.md",
"bytes": 3816,
"sha256_16": "0b7b390192dce873"
}
],
"quotes": [
{
"src": "skills/minimax-h3-text-to-video/references/template-and-examples.md",
"line": 8,
"text": "Create a [duration]-second [aspect-ratio] [format/genre] video as [one continuous shot / a concise sequence of N shots].\n\nSUBJECT AND SETTING\n…\nACTION AND TIMING\n[0–Xs]: [Opening composition and first action.]\n…\nCAMERA\n…\nVISUAL TREATMENT\n…\nAUDIO\n…\nCONSTRAINTS\n…",
"note": "★社区 T2V 模板是第四种习语:六节英文大写标题 + [0–Xs] 连字符窗。与官方 doc2 的 integrated_multimodal_description 三字段结构冲突 → 登记不采信(见 knowhow_v5 §冲突 C-4)"
},
{
"src": "skills/minimax-h3-reference-to-video/references/template-and-examples.md",
"line": 8,
"text": "REFERENCE ASSIGNMENTS\nUse Image 1 only for [identity/product/style/location/storyboard role]. Preserve [defining details]. Do not copy [unrelated influence].\nUse Video 1 only for [motion/camera/edit rhythm/performance role]. Adapt it to [new subject/space]. Do not copy [unrelated influence].\nUse Audio 1 for [voice/music/sound role]. Synchronize [mouth/breath/gesture/cuts/camera] to [specific moments].",
"note": "★社区「REFERENCE ASSIGNMENTS」节 = 每素材声明用途 + 反向排除(Do not copy …)。与官方 doc1:200「参考素材说明」同构,交叉印证 A 级做法;但它是英文自由文本,不是官方 <Subject N>/<Picture N> 标签体系 → 二者不等价,登记"
},
{
"src": "skills/minimax-h3-drama-producer/references/production-workflow.md",
"line": 5,
"text": "1. **Preflight** — inspect local capability and write the environment report.\n2. **Initialize** … 3. **Ingest** … 4. **Plan** — create the production brief, story/message beats, entity ledger, asset ledger, continuity plan, shot list, and budget.\n5. **Production approval** … 6. **Visual development** — create missing entity masters, scene masters, storyboard, and keyframes.\n7. **Visual-lock approval** … 8. **Shot generation** — route, prepare, submit, monitor, fetch, and score takes.\n9. **Post-production** … 10. **QC** … 11. **Delivery**",
"note": "★11 阶段生产工作流(含两个人工放行闸口),是「从分镜设计开始的全流程」最完整的社区参照",
"verbatim_lines": "5-15"
},
{
"src": "skills/minimax-h3-drama-producer/references/production-workflow.md",
"line": 23,
"text": "Design each shot as an atomic, controllable action. A short generative clip should normally contain two or three timed beats, not an entire sequence. Mark shots whose ending must bridge continuously into the next shot.",
"note": "单镜原子化 + 2-3 beats + 桥接镜标注(与 TFboy1 段内 2-3 分镜独立趋同)"
},
{
"src": "skills/minimax-h3-drama-producer/references/production-workflow.md",
"line": 25,
"text": "For each shot record:\n- narrative or message purpose;\n- exact duration and timeline position;\n- subjects, action order, performance, and screen direction;\n- framing, camera height, lens feel, movement, lighting, and texture;\n- dialogue, native sound, effects, music cue, and silence;\n- incoming and outgoing transition;\n- reference-to-role mapping and preservation constraints;\n- T2V, frame-to-video, first/last-frame, reference-to-video, or edit route;\n- ordinary or key-shot take budget;\n- expected end state and shot-specific hard gates.",
"note": "★单镜 10 项登记字段(含 路由选择 T2V/I2V/R2V/edit、take 预算、末态 hard gate)——比 v4 三要素多出「路由」「预算」「末态」三项,是 v5 全流程能力的直接素材",
"verbatim_lines": "25-36"
},
{
"src": "skills/minimax-h3-drama-producer/references/production-workflow.md",
"line": 40,
"text": "Create canonical masters only for recurring visual entities. Use the source type that matches the entity:\n- character sheet for people;\n- product sheet for product geometry and materials;\n- identity sheet for mascots, animals, and stylized subjects;\n- scene master for environment layout, fixed props, light direction, and axis.",
"note": "母版资产(entity/scene master)只为重复出现的视觉实体建 —— 一致性控制的最佳实践"
},
{
"src": "skills/minimax-h3-drama-producer/references/profiles/base-video/shot-patterns.md",
"line": 3,
"text": "- Establish the subject and environment before spatially complex action.\n- Use approved visual-entity and scene masters for recurring content.\n- Give a generative clip two or three timed beats at most.\n- Use frame control for approved compositions and reference control for identity, design, motion, camera, performance, or voice.\n- End every shot in a state that makes the next cut or bridge intentional.",
"note": "基础镜头模式 5 条(「End every shot in a state that makes the next cut intentional」= 衔接锚点社区表述)",
"verbatim_lines": "3-9"
},
{
"src": "skills/minimax-h3-adviser/references/prompt-language.md",
"line": 3,
"text": "Use this as a choice menu, not a checklist. Offer only 3–6 candidates relevant to the current unknown, explain their effect in plain language, and recommend one.",
"note": "★澄清/选项呈现纪律:候选只给 3–6 个 + 用平白语言解释效果 + 给一个推荐 —— 与 v4 的 clarify 轨直接相关,是社区给出的可量化澄清上限"
}
],
"used_for": [
"全流程 11 阶段",
"单镜 10 项登记",
"母版资产制度",
"澄清候选 3-6 个",
"T2V/R2V 模板习语登记为 C 级不采信"
]
},
{
"id": "s04",
"grade": "B",
"title": "GitHub 社区仓 · TFboy1/oh-my-minimaxh3-director(AI 视频导演流水线:自动分镜 → 工作流分配 → 批量生成 → 剪映拼合)",
"url": "https://github.com/TFboy1/oh-my-minimaxh3-director",
"default_branch": "master",
"description": "MiniMax H3 驱动的全自动 AI 视频导演流水线——自动分镜、工作流分配、参数确认、批量生成、剪映一键拼合。",
"crawl_time": "2026-08-28T17:20:00Z - 2026-08-28T17:27:00Z",
"tree_url": "https://api.github.com/repos/TFboy1/oh-my-minimaxh3-director/git/trees/master?recursive=1",
"tree_bytes": 12710,
"tree_sha256": "f7318f8717c6d221",
"tree_paths_total": 33,
"files_fetched": 8,
"fetch_via": "raw.githubusercontent.com 逐文件(8/8 HTTP 200,每请求间隔 2s);注意 default_branch=master,main 会 404",
"cache_root": "cache_v5/github/tfboy1/",
"files": [
{
"path": "README.md",
"bytes": 8760,
"sha256_16": "480372890e35c503"
},
{
"path": "SKILL.md",
"bytes": 16660,
"sha256_16": "0e404a5569a5bc9b"
},
{
"path": "agents/openai.yaml",
"bytes": 335,
"sha256_16": "63c610d509b429d9"
},
{
"path": "docs/README_EN.md",
"bytes": 8395,
"sha256_16": "309983adf490b0eb"
},
{
"path": "references/hybrid-example.md",
"bytes": 18421,
"sha256_16": "65b38bd9a242f13d"
},
{
"path": "references/storyboard-schema.md",
"bytes": 21396,
"sha256_16": "1774e9e428c96005"
},
{
"path": "references/wenwu-director.md",
"bytes": 17074,
"sha256_16": "155ad45dbf4663b9"
},
{
"path": "references/workflow-routing.md",
"bytes": 4598,
"sha256_16": "ffb8e8889fc82f28"
}
],
"quotes": [
{
"src": "SKILL.md",
"line": 101,
"text": "把剧本切段(默认每段 10 秒,5-15 秒可调),每段 2-3 个分镜,上下镜头之间写清衔接锚点(动作方向 / 视线 / 光线 / 道具 / 轮廓 / 声桥 / 受力),禁止无理由跳切。",
"note": "★段级切分 + 7 类衔接锚点枚举。段级 10s/5-15s 与官方 4-15s 上限一致;7 类锚点是社区归纳,无官方逐字出处 → C 级",
"verbatim_lines": "101-103"
},
{
"src": "SKILL.md",
"line": 107,
"text": "**先做提示词模式决策,再写提示词**:写提示词前必须判断简报复杂度,并把 `meta.prompt_mode` 与 `meta.prompt_mode_reason` 写入 storyboard.json,禁止静默走默认。",
"note": "★提示词模式显式决策 + 决策理由落盘(禁止静默默认)——可直接迁移为 v5 的模式选择判据做法"
},
{
"src": "SKILL.md",
"line": 132,
"text": "**分镜汇报(必须按秒段)**:分镜与提示词写完后,向用户汇报剧本时必须逐段给出**精确时间轴**——「几秒到几秒 → 什么画面 → 什么对白」,禁止用\"前面是开场、中间打斗、结尾反转\"这类宽泛概括。",
"note": "★分镜汇报硬格式(时间连续精确到秒、无空洞无重叠),与 storyboard-schema.md:319-347 的汇报格式表互证",
"verbatim": false,
"verbatim_lines": "132-137",
"reformat_note": "text 合并了 SKILL.md 的正文句与加粗片段;逐字以 cache :132-137 为准"
},
{
"src": "SKILL.md",
"line": 176,
"text": "提示词深度也做后端校验:六段齐全、镜头标记与时间码、hybrid 的 `constraints:` 块;`meta.strict_prompt_validation: true` 时不达标直接拒绝。",
"note": "提示词结构化后端校验 + strict 开关(可迁移为 v5 自检脚本思路)"
},
{
"src": "references/storyboard-schema.md",
"line": 7,
"text": "按叙事温度把故事切成段(段 = 一个 H3 视频单元,默认 10s,5-15s 可调),段内写 2-3 个分镜;**上下镜头之间必须有衔接**——每个镜头至少传递一个锚点(同一动作方向、同一视线目标、同一道光、同一个道具、同一种轮廓、同一段声音或同一股受力),禁止无理由跳切",
"note": "7 类衔接锚点完整表述",
"verbatim_lines": "7-11"
},
{
"src": "references/storyboard-schema.md",
"line": 13,
"text": "直接为每个主要角色生成**四视图**——三张不同视角的全身(正面 / 侧面 / 背面或 3/4 视角)+ 一张脸部特写;四视图必须锁定同一人物身份(脸、年龄、体型、发型、服装、饰品、惯用手)。",
"note": "★人物四视图制度(front/side/back/face 四文件命名约定见 :18-20)"
},
{
"src": "references/storyboard-schema.md",
"line": 389,
"text": "retention_analysis:\n<为什么观众记住本段,画面吸引力>",
"note": "★与官方冲突:官方 doc2:392 定义 retention_analysis 为「说明参考内容的保留、迁移或复用关系」+ 固定英文关系标记。该仓把字段语义误读为「观众记忆点」(字段名行在 :389,语义正文在 :390)→ 登记不采信(见 knowhow_v5 §冲突 C-3)",
"verbatim_lines": "389-390"
},
{
"src": "references/storyboard-schema.md",
"line": 387,
"text": "<10 秒内的剧情概要,包含 2-3 个分镜的时间线>",
"note": "与官方冲突:官方 doc2:360 要求 summary 开头用方括号任务类型标记([reference generation] 等),该仓缺失(字段名行 :386,语义正文 :387)→ 登记不采信"
},
{
"src": "references/storyboard-schema.md",
"line": 105,
"text": "`segments[].duration` 5-15 秒;`segments[].steps` 4-40;`segments[].seed` 0 到 2^31-1。",
"note": "段时长 5-15s(与官方 4-15s 上限相容但下限不同:社区取 5,官方允许 4)→ 登记差异"
},
{
"src": "references/storyboard-schema.md",
"line": 164,
"text": "每段可带 `shots` 数组(工业镜头卡,7 列严格):no / time / duration / shot_size / camera / content / sound / transition / technique / tags",
"note": "★7 列镜头卡 + 约束(:183-190:内容<=30字、声音列必填、镜号连续不跳号、时长累加=段长、半秒精度写 1.5\"、Flash Cut 0.3\"/0.5\")",
"verbatim": false,
"verbatim_lines": "164-190",
"reformat_note": "7 列镜头卡 text 是字段列举 + 约束合并(约束在 :183-190);逐字以 cache :164-190 为准"
},
{
"src": "references/storyboard-schema.md",
"line": 192,
"text": "### 景别 7 档(中文强制,禁止 ECU/CU 等英文缩写)\n| 景别 | 范围 | 主要用途 | 经典片例 |",
"note": "C 级自创:景别 7 档 + 机位库 11 + 运镜库 15 + 转场库 14 + 高级技法 8(:192-245)+ 剪辑节奏红线(ASL 偏离 >50% 诊断、σ<0.3 过平 >1.5 过抖,:270-278)。整套中文导演词汇无官方出处 → 只登记,不得写成 v5 硬判据"
},
{
"src": "references/storyboard-schema.md",
"line": 319,
"text": "## 分镜汇报格式(向用户呈现分镜时强制使用)\n汇报剧本时,每个镜头/段落必须包含:**起止时间(几秒到几秒)、画面内容、对白内容**。",
"note": "分镜汇报三要素硬格式"
},
{
"src": "references/storyboard-schema.md",
"line": 435,
"text": "## 段级生成原则(v1 固定)\n- 每段 = 一个 H3 视频,默认 10 秒,段内 2-3 个分镜由提示词驱动,段间用硬切接续、换新机位。\n- 提示词内禁止「参考上一段尾帧/续拍」,每段独立生成(衔接靠文字锚点,不靠画面引用)。",
"note": "★段独立生成 + 跨段衔接靠文字锚点而非画面引用 —— 这是多段拼长片时的关键工程约束(官方 15s 单次上限的必然推论,社区先给出可执行表述)",
"verbatim": false,
"verbatim_lines": "435-446",
"reformat_note": "段级生成原则原文是 5 条 bullet,text 摘录其中 3 条;逐字以 cache :435-446 为准"
}
],
"used_for": [
"段级切分 + 衔接锚点",
"分镜汇报硬格式",
"人物四视图",
"prompt_mode 显式决策",
"7 列镜头卡",
"retention_analysis/summary 语义冲突登记(不采信)",
"WenWu 导演词汇登记为 C 级(不采信)"
]
}
],
"reused_from_v4_no_refetch": {
"rule": "v4 已固化信源本轮一律禁复抓(未发任何请求);需要引用时逐字出处走 train/hunt/sources_v4.json 的 quotes 字段(含行号)",
"count": 25,
"registry": "train/hunt/sources_v4.json (s01-s25)",
"domains": [
"raw.githubusercontent.com/MiniMax-AI/MiniMax-H3 (官方仓 README / README.zh-CN / .claude/skills/h3-prompt-writing / skills/ 8 风格 skill / Ref2VA chat_template.json / scripts/readme 3 个可复现脚本)",
"api.github.com/repos/MiniMax-AI/MiniMax-H3 (git trees 全量路径)",
"platform.minimax.io(blocked,见 blocked_sources)"
],
"important_caveat": "v4 轮未把原文落盘(train/hunt 下无缓存目录,只有 sources_v4.json 注册表带 sha256_16 + 抽样 quotes),因此 v4 信源的「逐字回查」只能走注册表内 quotes,无法回查全文。本轮起 v5 已建立 cache_v5/ 全文落盘制度,后续轮次可全文回查。",
"load_bearing_for_v5": [
{
"v4_id": "s05",
"claim": "Output duration 4–15 seconds(README.md:73)",
"now_superseded_by": "s01 (doc1:35) A 级一手「输出时长 4-15秒」"
},
{
"v4_id": "s02",
"claim": "base-en.txt:86 首镜不写时间戳 + strictly increasing cut time",
"now_superseded_by": "s02 (doc2:107/:111) A 级一手逐字"
},
{
"v4_id": "s02",
"claim": "base-en.txt:94-96 运镜三维度",
"now_superseded_by": "s02 (doc2:121) A 级一手逐字"
},
{
"v4_id": "s02",
"claim": "base-en.txt:146 画面文字双引号",
"now_superseded_by": "s02 (doc2:154) A 级一手逐字"
},
{
"v4_id": "s03",
"claim": "ref-en.txt:73 <Picture 3> is a storyboard reference for [Shot 1] and [Shot 2]",
"now_superseded_by": "s02 (doc2:310) A 级一手逐字"
},
{
"v4_id": "s13",
"claim": "co-op 模板六时间窗 [0秒–2秒]…[10秒–15秒](master v4 案例官方同型)",
"now_superseded_by": "s01 (doc1:97) A 级一手同型案例(六个窗逐一相同)"
},
{
"v4_id": "s19",
"claim": "handdrawn SKILL.md:13 跟随用户语言(习语乙语言规则)",
"status": "仍为 v4 镜像出处,未被本轮取代;与 s02 (doc2:242) 构成官方内部冲突,见 knowhow_v5 §冲突 C-2"
}
]
},
"negative_findings": [
{
"claim": "两篇飞书文档均无 <whiteboard token/> 受限渲染",
"checked_by": "对两篇缓存 md 全文 grep '<whiteboard' 计数",
"result": "doc1=0、doc2=0 命中 → 本轮两篇全文完整可读,无 lark-cli 白板受限项损失(--doc-format markdown 模式)"
},
{
"claim": "飞书侧抓取是否需要按 block 分段",
"checked_by": "对比 outline 输出(doc1 51 个标题节点 / doc2 26 个)与 --doc-format markdown 全文返回",
"result": "两篇都一次取全(147058B / 41607B),未触发 --scope section 分段需求;`--scope section` 本轮未使用"
},
{
"claim": "doc2 是否包含 [N秒–M秒] 窗式时间写法",
"checked_by": "通读 doc2 全文 649 行",
"result": "零命中。doc2 的时间表达只有 [Shot N](首镜无时间戳)与 `At MM:SS.mmm`(:114/:455)。窗式写法只出现在 doc1(:97/:200/:237)→ 「窗式 vs [Shot N]+MM:SS.mmm」是官方一手内部差异,不是镜像讹误(见 knowhow_v5 §冲突 C-1)"
},
{
"claim": "doc2 是否给出 negative prompt / 负向约束的独立字段",
"checked_by": "通读 doc2 全文 + grep 'negative|禁|不得|不要'",
"result": "doc2 六段结构中没有独立负向约束段;负向信息只允许落在 overall_soundscape/non_diegetic_music 写 N/A(:162/:170)与 <d> 忠实转写(:575)。独立的负向块(no hard cuts…)只出现在 v4 s13 co-op 模板 :133 与 doc1:200 的「不想要」栏 → 负向约束的官方归属是「画面过程说明的不想要栏」,不是第六段"
},
{
"claim": "raw.githubusercontent.com / github.com 的 robots.txt 是否允许本轮抓取",
"checked_by": "curl --retry 1 --max-time 40 抓 /robots.txt(两域)",
"result": "均超时未取得 → 本轮无法按 SOP 逐字核对 robots 规则,属「无法复检」。处置:继续推进但保持最低侵入(只抓公开仓文本文件、无并发、间隔 >= 2s、不绕登录),并把此项登记为未决。下轮复检时应先抓 robots 再抓正文"
},
{
"claim": "chiphoton 仓是否自带官方 <Subject N>/<Picture N> 标签体系",
"checked_by": "通读其 4 个 template-and-examples.md + adviser/drama-producer SKILL",
"result": "未发现 <Subject N>/<Picture N>/<Audio N> 标签用法;其 reference 模板用英文自由文本 REFERENCE ASSIGNMENTS 表达同类意图 → 与官方标签体系不等价,不能把社区模板当成官方全能参考格式的范例"
},
{
"claim": "两个社区仓对 15s 上限的处理是否一致",
"checked_by": "对比 TFboy1 storyboard-schema:105 (duration 5-15) 与 chiphoton production-workflow(未给数字,用 take budget 管控)",
"result": "两者都不突破官方 4-15s;TFboy1 显式把「每段 = 一个 H3 视频」钉在 5-15s 并用多段拼长片(:435-446),chiphoton 用 11 阶段流水线 + take budget。与官方 doc1:35 一致"
}
],
"blocked_sources": [
{
"url": "https://platform.minimax.io/docs/api-reference/video-generation-v2-create",
"reason": "沿用 v3/v4 状态(WebFetch 域安全校验被拒 + curl 超时),本轮未重试;本轮问题集不涉及开放平台 schema",
"impact": "开放平台请求 schema 继续未核实。注意:doc1:15 给出了官方开放平台文档新地址 https://platform.minimaxi.com/docs/api-reference/video-generation-v2-create(.minimaxi.com 非 .minimax.io),下轮应改抓该域名"
},
{
"url": "https://raw.githubusercontent.com/robots.txt 与 https://github.com/robots.txt",
"reason": "本轮抓取超时,robots 规则未核对(见 negative_findings)",
"impact": "礼貌抓取约束按保守执行;登记为未决,不视为失败"
}
],
"verification_of_no_overwrite": {
"rule": "v1-v4 产物一律零改动",
"checked_at_utc": "2026-08-28T17:30:00Z",
"previously_observed_bytes": {
"sources_v4.json": 33147,
"knowhow_v4.md": 33756,
"feishu_access_status.md": 1854
},
"method": "本轮全程只新建文件(cache_v5/ 34 文件 + knowhow_v5.md + sources_v5.json + crawl_config_v5.json),未对任何既有文件执行写/删/rename;收工后复测字节数见下",
"post_check": "见本文件 bottom 的 post_check 字段"
},
"post_check": {
"note": "开工基线(写自 sources_v5.json 正文 previously_observed_bytes)与收工复测逐一致;mtime 全部早于本轮开工时刻 2026-08-28T17:05Z,即本轮零改动",
"sources_v4.json": {
"bytes": 33147,
"mtime_utc": "2026-08-27T15:03:54.926Z"
},
"knowhow_v4.md": {
"bytes": 33756,
"mtime_utc": "2026-08-27T15:01:09.862Z"
},
"feishu_access_status.md": {
"bytes": 1854,
"mtime_utc": "2026-08-25T10:05:19.942Z"
},
"master_input_storyboard_case_v4.md": {
"bytes": 3157,
"mtime_utc": "2026-08-27T14:47:20.814Z"
},
"sources_v3.json": {
"bytes": 19675,
"mtime_utc": "2026-08-27T06:57:39.124Z"
},
"knowhow_v3.md": {
"bytes": 35253,
"mtime_utc": "2026-08-27T06:56:04.029Z"
},
"sources.json": {
"bytes": 8049,
"mtime_utc": "2026-08-25T10:05:19.942Z"
},
"knowhow_v1.md": {
"bytes": 18763,
"mtime_utc": "2026-08-25T10:05:19.942Z"
},
"v4_baseline_bytes": {
"sources_v4.json": 33147,
"knowhow_v4.md": 33756,
"feishu_access_status.md": 1854
},
"verdict": "PASS(v4 三文件字节数与基线一致且未被触碰)"
},
"quote_text_policy": {
"rule": "quotes[].text 是「定位型摘录」。凡 `verbatim:false` 者表示 text 为表格/列表的紧凑转写(省略以 … 标示),派单方逐字回查必须以 `verbatim_lines` 指向的 cache_v5/ 原文为准;凡未标 verbatim 者默认 true,text 应与 cache 原文逐字一致(仅空白/换行差异)。",
"verified_by": "offset-to-line 全文审计:82 条 quote 中 71 条 text 与 cache 逐字一致且行号精确(line-off 0);11 条表格/列表转写已标 verbatim:false 并给 verbatim_lines(span 校验全过)",
"audit_date_utc": "2026-08-28"
}
}