爆款仿写 Workflow

Use this skill when the user wants an agent to turn a topic, account role, or viral reference into a spoken short video.

Read these first when needed:

```
../luma-shared/SKILL.md
```
for common project, artifact, auth, and failure rules.
```
../luma-content-research/SKILL.md
```
for topic search and keyword tables.
```
../luma-material/SKILL.md
```
for local material groups and PIP matching.
```
../luma-digital-human/SKILL.md
```
for voice, TTS, avatar, and lip-sync.
```
../luma-subtitle/SKILL.md
```
for subtitle rendering.

When To Use

The user asks for 爆款仿写, 对标视频, 口播短视频, 种草视频, or a complete video production run.
The user wants all intermediate files so the agent can inspect and iterate.
The expected output is a produced video plus cover, not just a script.

Do not use this workflow when the user only asks for one atomic operation such as TTS, subtitle, or material search.

文案是根基：严禁跳过 Research

爆款仿写的核心是"仿写"，不是"原创"。文案的选题、结构、节奏必须基于真实爆款数据，绝不能凭 AI 自己拍脑袋编。

为什么 Step 0 (Research) 不能跳过

没有数据支撑的文案是盲猜。你不知道什么选题正在爆、什么结构观众买单、什么钩子点击率高。
仿写的前提是有对标。Step 0 输出的是：热门关键词、对标视频链接、爆款标题、点赞量、口播/非口播分类。这些信息决定了 Step 1 写什么。
跳过 Step 0 直接自己写 = 把"仿写"变成了"盲写"。后面 TTS、lipsync、字幕、BGM 做得再好，方向错了全白费。

Agent 执行规则（强制）

Step 0 Research 不可跳过。 不管用户有没有明确要求，必须先跑
```
research run
```
。如果用户说"随便写一个"，你要拒绝，告诉他需要数据支撑选题。
Step 1 必须基于精选对标视频转写。 Research 只负责发现候选爆款，不能直接当 source script。必须先筛选候选，只选择最多 3 个高讨论度、高点赞、强相关、且大概率是口播/讲述型的视频下载并 ASR。source script 必须能追溯到这些精选转写结果。
禁止 AI 自己编 source script。 不允许只看标题、点赞量、关键词就凭空写
```
source_script.txt
```
。如果对标视频无法下载或 ASR 失败，必须换参考视频；仍无法获得转写时，暂停并向用户说明无法完成"仿写"依据。
Research 结果要展示给用户。 跑完 Step 0 后，列出找到的关键词、Top 3 对标视频（标题+点赞量），让用户知道文案的选题依据是什么。

Standard Files

```
step0_content_research.json
```
```
step0_content_research.csv
```
```
step0_keywords.json
```
```
step0_keywords.csv
```
```
references/ref_01.mp4
```
```
references/ref_01_asr.json
```
```
references/ref_02.mp4
```
```
references/ref_02_asr.json
```
```
source_reference_bundle.md
```
```
source_script.txt
```
```
step1_rewrite.json
```
```
transcript.txt
```
```
step2_tts.wav
```
```
step3_lipsync.mp4
```
```
step4_segments.json
```
```
step4_scene_units.json
```
```
step4_materials_enriched.json
```
```
step4_material_matches.json
```
```
step4_picture_in_picture_plan.json
```
```
step4_picture_in_picture.mp4
```
```
step5_subtitle.mp4
```
```
step6_bgm.mp4
```
```
step7_covers/cover_manifest.json
```
```
step7_covers/cover_01.jpg
```

Flow

Create or select a project:

bash

luma-cli project create viral-remix
luma-cli project use viral-remix

Research references:
bash
```
luma-cli research run --role "<role_or_topic>" --mode precise --date-range 7d --output step0_content_research.json
luma-cli research export --input step0_content_research.json --output step0_content_research.csv
luma-cli research keywords --input step0_content_research.json --output step0_keywords.json --csv step0_keywords.csv
```
Inspect the JSON/CSV and build a shortlist before spending ASR credits. Do not download every result.

Selection rules:
- Choose at most 3 references.
- Prefer videos with high likes, strong discussion potential, clear controversy/curiosity, and close fit to the user's topic/persona.
- Prefer
```
content_type=口播
```
  or videos whose title/description suggests spoken explanation, opinion, review, teaching, story, or analysis.
- Skip pure music/card-point edits, scenery montages, product-only showcases, dance clips, meme clips, or any video likely to have little reusable spoken copy.
- If all high-like videos are non-spoken, pick a topic cluster only and rerun research with a more口播-oriented role/query; do not ASR weak references just to fill the quota.
- Record the shortlist and rejection reasons in
```
source_reference_bundle.md
```
  before downloading.
Download and transcribe the chosen references:
bash
```
mkdir -p references
luma-cli --json social download "<reference_1_link>" --output references/ref_01.mp4
luma-cli asr references/ref_01.mp4 --language zh --output references/ref_01_asr.json
luma-cli --json social download "<reference_2_link>" --output references/ref_02.mp4
luma-cli asr references/ref_02.mp4 --language zh --output references/ref_02_asr.json
```
If using a third reference, save it as
```
references/ref_03.mp4
```
and
```
references/ref_03_asr.json
```
.
```
asr
```
accepts video directly, so a separate local audio-extraction step is not required unless ASR fails on the video file.
After each ASR result, check whether the transcript is actually useful. If the transcript is empty, mostly music/noise, too short to reveal structure, or unrelated to the title, discard that reference and choose another shortlisted video. Stop once 1-3 strong transcripts are available; do not keep spending ASR on weak candidates.
Build the source material for rewrite:
- Read each
```
references/ref_XX_asr.json
```
  .
- Extract the original transcript, hook, argument structure, emotional turn, punchlines, and CTA.
- Write
```
source_reference_bundle.md
```
  with the selected references, original titles/links, transcript excerpts, and the reason each reference is worth copying.
- Write
```
source_script.txt
```
  as a grounded source brief: include the user's persona/positioning, the 2-3 reference viewpoints to fuse, reusable structures/hooks from the transcripts, and explicit constraints. Do not invent claims that are absent from the reference transcripts or user brief.
Rewrite the grounded source script:
bash
```
luma-cli script rewrite --input source_script.txt --length short --output step1_rewrite.json
```
Save the rewritten text as
```
transcript.txt
```
for later subtitle steps (avoids redundant ASR).

Generate speech from the rewritten text:

bash

luma-cli --json tts --file transcript.txt --voice 男声3 --speech-rate 1.1 --output step2_tts.wav

The

--json

flag outputs

audio_object_key

which can be passed directly to lipsync, avoiding a redundant upload.

Generate digital-human video (use

--audio-key

to reference the cloud audio directly):

bash

luma-cli lipsync --avatar 数字人男 --audio-key <audio_object_key> --random-start --output step3_lipsync.mp4

--audio-key

is omitted, lipsync falls back to the project's

latest_tts_key

, then to

--audio

file upload.

Segment text and build scene units:

bash

luma-cli subtitle transcript.txt --text --segments-output step4_segments.json --no-effects --no-highlight
luma-cli pip scene --segments step4_segments.json --output step4_scene_units.json

Prepare and match local PIP materials:

bash

luma-cli material group describe vlm_ai --output step4_materials_enriched.json
luma-cli pip match --scenes step4_scene_units.json --materials step4_materials_enriched.json --mode auto --output step4_material_matches.json

Plan and render PIP:

bash

luma-cli pip plan --segments step4_segments.json --materials step4_materials_enriched.json --match-mode auto --output step4_picture_in_picture_plan.json
luma-cli pip render step3_lipsync.mp4 --plan step4_picture_in_picture_plan.json --output step4_picture_in_picture.mp4

If no insert is matched, continue with

step3_lipsync.mp4

as the subtitle input.

Add subtitles (uses
```
--transcript
```
to skip ASR since we already have the exact script):

bash

luma-cli subtitle step4_picture_in_picture.mp4 --transcript transcript.txt --output step5_subtitle.mp4

Add BGM:

bash

luma-cli bgm mix step5_subtitle.mp4 --output step6_bgm.mp4

Create a cover:
bash
```
luma-cli cover generate step4_picture_in_picture.mp4 --title "<cover_title>" --subtitle "<cover_subtitle>" --count 12 --output-dir step7_covers
```
Cover source rule: use a clean visual video before burned subtitles and BGM. Prefer
```
step4_picture_in_picture.mp4
```
; if PIP was skipped, use
```
step3_lipsync.mp4
```
. Never use
```
step5_subtitle.mp4
```
or
```
step6_bgm.mp4
```
as the cover source, because burned subtitles will become part of the cover background.

Agent Rules

Keep every intermediate file; do not collapse the flow into one hidden step.
```
source_script.txt
```
is not the final script and not AI-written from memory. It is the grounded rewrite brief produced from downloaded reference transcripts plus the user's persona/angle.
If
```
step0_content_research.json
```
only has titles/links and no transcript, do not proceed to rewrite until reference videos have been downloaded and ASR has produced usable text.
ASR is an expensive validation step, not a bulk-processing step. Never ASR more than 3 reference videos for one remix unless the user explicitly approves it.
Use the rewritten script as the single source for TTS, segmentation, subtitles, and cover text extraction.
Covers must use clean visual frames, not videos that already contain burned subtitles.
If the material library does not fit the script, skip PIP instead of forcing weak matches.
Use
```
project artifact list
```
before resuming or rerunning a partial workflow.
Report exact output paths for all generated files.

luma-workflow-viral-remix

NPX Install

Tags

SKILL.md Content

爆款仿写 Workflow

When To Use

文案是根基：严禁跳过 Research

为什么 Step 0 (Research) 不能跳过

Agent 执行规则（强制）

Standard Files

Flow

Agent Rules