watch

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

The CLI is self-describing and ships its own API reference. Use

dapi --help

dapi media --help

, and

dapi media <command> --help

to enumerate every command, argument, and option, and treat live help as authoritative rather than working from memory. If

dapi

is unavailable, read installation.md.

该CLI工具自带说明文档和API参考。使用

dapi --help

、

dapi media --help

和

dapi media <command> --help

来查看所有命令、参数和选项，应以实时帮助信息为准，而非凭记忆操作。如果

dapi

未安装，请阅读installation.md。

Watching footage

解析音视频素材

How to understand source material. Inspect only the modalities the question turns on — speech, action, music, graphics, or atmosphere may lead, so there is no fixed priority. Sample the picture against what the audio tells you.

Always probe first.
```
dapi media probe <id|path>
```
reports the container and its tracks, telling you up front whether the file has a video track, an audio track, or both. Everything after branches on that.
Get the lay of the land. Render a
```
dapi media waveform
```
(audio) and a
```
dapi media filmstrip
```
(video) for a fast, cheap overview of where the loud and quiet stretches fall, and where the visual scene changes are. A filmstrip shows coarse structure and scene state, not crop, framing, readability, or an exact cut frame.
Listen to the audio. Run
```
dapi media listen
```
with a prompt tailored to the context (what you actually need to know), and explicitly ask the model to include timestamps in its answer. See media-listen.md for prompt patterns.
Transcribe speech. For speech,
```
dapi media transcribe
```
prints the full transcript with word-level start/end times directly — read any segment straight from it.
Sample the video against the audio. Use
```
dapi media grab
```
to pull frames. When the audio has already pointed you at specific moments, feed those timestamps straight in from the transcript or listen output, e.g.
```
-t '00:32' '00:45' ...
```
. When you need a visual pass without such cues, reach for
```
--auto
```
: it scans the footage and keeps only the frames where the picture settles into a new visual state, dropping near-duplicates.

如何解析源素材。仅针对问题涉及的维度进行检查——语音、动作、音乐、画面或氛围都可能是重点，因此没有固定优先级。结合音频信息来比对画面内容。

务必先探测文件信息。
```
dapi media probe <id|path>
```
会报告文件容器及其轨道信息，让你提前了解该文件是否包含视频轨道、音频轨道，或是两者都有。后续所有操作都以此为基础展开。
快速掌握素材整体情况。生成
```
dapi media waveform
```
（音频波形图）和
```
dapi media filmstrip
```
（视频分镜图），快速低成本地了解音频的高低音段落分布，以及视觉场景的变化位置。分镜图仅展示大致结构和场景状态，不包含裁剪、构图、可读性或精确的剪辑帧信息。
分析音频内容。运行
```
dapi media listen
```
并结合上下文定制提示词（明确你实际需要了解的内容），并要求模型在回答中明确包含时间戳。提示词模板可参考media-listen.md。
转录语音内容。针对语音内容，
```
dapi media transcribe
```
会直接输出包含逐词起止时间的完整转录文本——可直接从中读取任意片段。
结合音频信息采样视频画面。使用
```
dapi media grab
```
提取帧画面。当音频已经指向特定时刻时，直接从转录文本或
```
listen
```
的输出中获取时间戳传入，例如
```
-t '00:32' '00:45' ...
```
。如果没有此类线索需要全面扫描画面，可使用
```
--auto
```
参数：它会扫描素材并仅保留画面切换到新视觉状态的帧，去除近似重复的帧。

Matching depth to the question

根据问题深度匹配操作

Read only as much of the footage as the answer requires — each pass costs time, and

listen

costs credits.

A duration or format question ends at
```
probe
```
.
"Where is the quiet part" or "how is it paced" is usually answered by the waveform and filmstrip alone.
A question about what was said resolves fastest through
```
transcribe
```
; quote the transcript and its times directly.
Questions about non-speech audio (music, tone, sound events, speaker identity) are what
```
listen
```
is for.
Only questions about what is seen need frames — and the audio pass usually tells you which moments to grab, so grab those instead of scanning blind.
For open-ended questions ("summarize this", "what happens here"), combine passes: structure from waveform + filmstrip, content from transcript or listen, then frames at the salient moments to confirm what the picture shows.

仅解析回答问题所需的素材内容——每次解析都会耗时，且

listen

操作会消耗积分。

关于时长或格式的问题，仅需执行
```
probe
```
操作即可得到答案。
“安静段落在哪里”或“节奏如何”这类问题通常仅通过波形图和分镜图就能回答。
关于语音内容的问题，通过
```
transcribe
```
操作能最快得到答案；直接引用转录文本及其时间戳即可。
关于非语音音频（音乐、语气、声音事件、说话人身份）的问题，适合使用
```
listen
```
操作。
只有涉及视觉内容的问题才需要提取帧画面——而音频分析通常会告诉你需要抓取哪些时刻的画面，因此直接抓取这些时刻即可，无需盲目扫描。
对于开放式问题（如“总结这段内容”“这里发生了什么”），需结合多种操作：通过波形图+分镜图了解结构，通过转录文本或
```
listen
```
获取内容，然后在关键时刻提取帧画面来确认视觉内容。

Answering

回答要点

Ground every claim in something you actually saw or heard — a transcript line, a listen answer, a grabbed frame. If the evidence is ambiguous, say so rather than smoothing over it.
Anchor answers to the timeline. Give timestamps as
```
MM:SS
```
(or
```
HH:MM:SS
```
for long footage) so the user can jump straight to the moment; for a scene or segment, give its start and end.
When asked to find a scene or moment, return the timestamp range plus a one-line description of what identifies it, so the user can confirm it is the right one.
Summaries follow the footage's own structure: what happens, in order, with the timestamps where each part begins. Length matches what the user asked for, not what the footage contains.
This skill only reads footage. When the user wants the footage changed — cut, composed, captioned, exported — that is the
```
editor
```
skill's job.

所有结论都必须基于你实际看到或听到的内容——转录文本、
```
listen
```
的回答、抓取的帧画面。如果证据模糊不清，需明确说明，而非含糊带过。
回答需锚定时间线。时间戳格式使用
```
MM:SS
```
（长素材使用
```
HH:MM:SS
```
），方便用户直接跳转到对应时刻；对于场景或片段，需给出起止时间。
当用户要求查找某个场景或片段时，返回时间戳范围并附上一行描述说明该场景的识别特征，以便用户确认是否为目标内容。
摘要需遵循素材自身的结构：按顺序描述事件内容，并标注每个部分的开始时间。摘要长度需符合用户要求，而非取决于素材本身的长度。
本技能仅用于解析素材。当用户需要修改素材——剪辑、合成、添加字幕、导出——则属于
```
editor
```
技能的范畴。