Back to Details

descript

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

descript

descript

The talk-content editing tool skill — write the edit in the transcript, Overdub with consent, refine the sound, dress the visuals, and ship the cuts. The agent plans (and can drive Underlord via API/MCP where connected); the human verifies by ear and approves; WoopSocial publishes the exports. (Ships with 
tools/integrations/descript.md
.)
口播内容编辑工具 skill——在转录文本中完成剪辑、经同意后使用 Overdub、精修声音、包装视觉效果并交付剪辑成品。Agent 负责规划（已连接时可通过 API/MCP 驱动 Underlord）；人类通过试听进行验证并批准；WoopSocial 发布导出内容。（随附 
tools/integrations/descript.md
。）

The POV: the transcript is the timeline — decide in text, verify by ear

观点：转录文本就是时间线——以文本决策，以试听验证

For dialogue-heavy content, editing the transcript beats scrubbing a timeline: delete the sentence, the clip disappears; move the paragraph, the footage follows — reviews report ~60–70% editing-time cuts for talk content. But the paradigm has two sharp edges the top 1% respect. (1) The voice spine: Overdub's consent-verified, own-voice-only design is the model, not an obstacle — it exists so nobody types words into someone else's mouth; and the craft truth is it shines on flubbed words, not paragraphs (long Overdub drifts synthetic — re-record those). (2) The meaning spine: text-editing makes it dangerously easy to rearrange a guest into saying something they didn't — concision yes, meaning-flips never, and the human owns the final cut of anyone else's words. Operationally: the accuracy pass is mandatory (transcript errors become wrong edits AND wrong captions), and since the Sept 2025 overhaul, the workflow must be credit-aware — media minutes count everything you import, and formerly-unlimited AI features are metered.
对于对话密集型内容，编辑转录文本比反复拖动时间线更高效：删除一句话，相应片段随之消失；移动一个段落，画面也会随之移动——相关评测称，对于口播内容，这种方式可减少约 60–70% 的编辑时间。但这一范式也有两处必须谨慎对待的利刃，顶尖 1% 的实践者都会正视它们。(1) 声音主线：Overdub 采用经同意验证、仅限自有声音的设计，这正是应当效仿的范例，而不是障碍——它是为了避免让别人的嘴说出并不属于他们的话；从制作角度看，它最擅长修正说错的词，而不是整段内容（长篇 Overdub 会产生合成感——这类内容应重录）。(2) 意义主线：文本编辑很容易造成危险结果：将嘉宾的话重排成他们并未说过的话——可以精简，绝不改变原意，而任何他人言论的最终剪辑权都应由人来掌握。操作上：准确性校对是必做步骤（转录错误会变成错误剪辑和错误字幕），而且自 2025 年 9 月改版以来，工作流必须关注 credits 消耗——媒体分钟数会计入所有导入的素材，以前不限量使用的 AI 功能现在也改为按量计量。

Read these first

首先读取

  1. The recording's content skill — podcast-and-audiograms / youtube-long-form / educational-content.
  2. brand-profile + voice-builder (written outputs) + design-and-templates (captions/layout).
  1. 该录音对应的内容 skill——podcast-and-audiograms / youtube-long-form / educational-content。
  2. brand-profile + voice-builder（书面输出）+ design-and-templates（字幕和布局）。

The framework: WORDS

框架：WORDS

(Depth: 
references/the-words-framework.md
.)
  • W — Write the edit: accuracy pass first; then cut tangents/bad takes, one-step filler+silence removal (Underlord), restructure by moving paragraphs — decide in text, verify by ear.
  • O — Overdub with consent: own-voice-only, consent-verified; single words/short phrases (paragraphs = re-record); vocabulary/credit limits; disclose synthetic speech where required.
  • R — Refine the sound: Studio Sound once per source (credits; cleaner audio also improves the transcript); level speakers; extreme noise is a re-record, not a rescue.
  • D — Dress the visuals: Automatic Multicam (record separate tracks on purpose), captions from the corrected transcript, human-approved B-roll, Eye Contact used honestly; beat-sync/color route elsewhere.
  • S — Ship the cuts: one transcript → the episode + clips (Underlord flags, the human picks fairly) + show notes + chapters + a text post; route onward and publish via WoopSocial.
（深入内容：
references/the-words-framework.md
。）
  • **W — 以文本完成剪辑：**先进行准确性校对；然后剪掉离题内容和失败录制，通过 Underlord 一步移除填充词和静音；通过移动段落重组结构——在文本中决策，通过试听验证。
  • **O — 经同意使用 Overdub：**仅限自有声音，并须经同意验证；只用于单个词或短语（整段内容应重录）；受词汇和 credits 限制；在要求披露时说明使用了合成语音。
  • **R — 精修声音：**每个源文件运行一次 Studio Sound（消耗 credits；更干净的音频也能改善转录）；统一各说话者的音量；噪声严重时应重录，而不是强行修复。
  • **D — 包装视觉效果：**Automatic Multicam（有意分别录制多条轨道）、根据校正后的转录生成字幕、使用经人工审核的 B-roll、如实使用 Eye Contact；节拍同步和色彩处理转至其他流程。
  • **S — 交付剪辑成品：**一份转录文本 → 单集节目 + 片段（由 Underlord 标出候选内容，人工公平挑选）+ 节目说明 + 章节 + 文字帖文；继续分发并通过 WoopSocial 发布。

The reality (verify-quarterly)

实际情况（每季度核实）

2026 Descript: Underlord (agentic co-editor — filler/silence in one step, bad-take flags, B-roll suggestions, clips, show notes) now triggerable via the 2026 public API (open beta) incl. MCP connections; Overdub (~24–48h training; source-audio requirements have varied — verify); Studio Sound (~10 credits/use); Automatic Multicam; Eye Contact; ~92–95% transcription accuracy on clean audio, ~75–85% with noise/accents/jargon; ~23 languages; SOC 2 Type II; cloud-dependent (no offline). Pricing: the Sept 2025 overhaul moved to media minutes + AI-credit metering of formerly-unlimited features; documented bill-shock and no mid-cycle proration (G2, attributed); tier figures conflict across sources — verify in-app. Full detail: 
references/descript-2026-reality.md
. The weekly loop, credit-aware checklist, the Overdub decision table, the interview-integrity checklist, and two worked examples: 
references/workflows-and-templates.md
.
2026 年 Descript：Underlord（Agent 式协同编辑工具——一步移除填充词和静音、标记失败片段、建议 B-roll、生成片段和节目说明）现已可通过 2026 年公共 API（公测）触发，包括 MCP 连接；Overdub（训练约 24–48 小时；源音频要求可能有所变化——需核实）；Studio Sound（每次约消耗 10 credits）；Automatic Multicam；Eye Contact；干净音频的转录准确率约为 92–95%，有噪声、口音或行业术语时约为 75–85%；支持约 23 种语言；SOC 2 Type II；依赖云端（不支持离线使用）。定价方面，2025 年 9 月的改版改为按媒体分钟数 + AI credits 计量计费，原先不限量使用的功能也纳入计量；G2 上已有署名反馈记录了费用意外上涨和周期中途不按比例折算的问题；不同来源给出的套餐数字相互冲突——请在应用内核实。完整详情见：
references/descript-2026-reality.md
。每周工作循环、credits 用量检查清单、Overdub 决策表、访谈完整性检查清单以及两个实战示例见：
references/workflows-and-templates.md
。

Honest scope (never violate)

真实能力边界（绝不逾越）

  • The agent plans the edit and can trigger Underlord/media actions via the API/MCP where connected (exact human steps otherwise — no pretended automation); the human verifies by ear (pacing/tone/fairness don't live in text) and approves — the agent never fabricates "that cut sounds great." WoopSocial publishes the finished exports; it does not edit media; podcast RSS distribution is separate (human; podcast-and-audiograms).
  • Voice spine: own-voice-only cloning; never a guest/competitor/public figure; never fabricated words in a real mouth; AI-disclosure for synthetic speech where required (EU AI Act; C2PA). Meaning spine: interview edits preserve meaning + clip context; approval offered on significant edits. Never fabricate tiers, credits, or metrics — verify in-app. (Full scope: 
    references/scope-and-connections.md
    .)
  • Agent 负责规划剪辑，并可在已连接时通过 API/MCP 触发 Underlord 或媒体操作（未连接时提供准确的人工操作步骤——不假装已经实现自动化）；人类通过试听进行验证（节奏、语气和公平性无法仅靠文本体现）并予以批准——Agent 绝不编造“这段剪得真棒”之类的评价。WoopSocial 发布最终导出内容，但它不负责编辑媒体；播客 RSS 分发是独立流程（由人工负责；podcast-and-audiograms）。
  • **声音主线：**仅限克隆自有声音；绝不用于嘉宾、竞争对手或公众人物；绝不让真实人物之口说出编造的话；在法规要求时对合成语音进行 AI 披露（EU AI Act；C2PA）。**意义主线：**访谈剪辑必须保留原意和片段上下文；重大编辑应提供审批选项。绝不编造套餐、credits 或指标——请在应用内核实。（完整范围：
    references/scope-and-connections.md
    。）

Distinct from its siblings (route correctly)

与同类 skill 的区别（正确分流）

descript (this) = text-based talk-content editing · capcut = beat-synced visual short-form (the hybrid: master here, style cuts there) · captions-and-clipping / opus-clip = clip selection at scale (this feeds them the master) · podcast-and-audiograms = the strategy this tool serves · ai-voiceover (elevenlabs) = dedicated TTS/narration (Overdub = own-voice corrections) · talking-head-and-piece-to-camera = the performance (Eye Contact patches a read, doesn't replace delivery) · youtube-long-form = the structure the recording follows.
descript（本 skill） = 基于文本的口播内容编辑 · capcut = 按节拍同步的视觉短内容（混合流程：母版在这里，风格化剪辑在 CapCut 中完成） · captions-and-clipping / opus-clip = 规模化筛选片段（descript 为其提供母版） · podcast-and-audiograms = 本工具所服务的策略 · ai-voiceover (elevenlabs) = 专用 TTS/旁白（Overdub = 自有声音纠错） · talking-head-and-piece-to-camera = 表演形式（Eye Contact 可以修饰一次朗读，但不能替代表达表现） · youtube-long-form = 录音所遵循的结构。

Where this connects

连接关系

Reads first: podcast-and-audiograms / youtube-long-form + brand-profile/voice-builder + design-and-templates. Feeds: captions-and-clipping / opus-clip (the master), capcut (short-form styling), text-post-and-microblog (the written cut), email-and-newsletter (show notes), youtube-publishing-and-metadata. Publishes via: exports → scheduling-and-queue → WoopSocial (social); RSS host (podcast — human). Tool file: 
tools/integrations/descript.md
. Measure with: native + analytics-and-reporting on listen-through/watch-through + clips — never fabricated.
首先读取：podcast-and-audiograms / youtube-long-form + brand-profile/voice-builder + design-and-templates。输出至：captions-and-clipping / opus-clip（母版）、capcut（短内容风格化）、text-post-and-microblog（文字剪辑）、email-and-newsletter（节目说明）、youtube-publishing-and-metadata。发布路径：导出内容 → scheduling-and-queue → WoopSocial（社交媒体）；RSS 托管（播客——由人工负责）。工具文件：
tools/integrations/descript.md
。衡量方式：使用原生分析及 analytics-and-reporting 评估完听、完整观看和片段表现——绝不编造数据。

Definition of done

完成定义

A talk-content edit made at the speed of text and verified by ear: transcript corrected first (names/jargon — errors become wrong edits and captions), tangents/bad takes cut and filler+silences removed in batched Underlord passes, restructure done in text and the result listened through; flubs fixed with consent-verified own-voice Overdub at word/phrase length only (paragraphs re-recorded; synthetic speech disclosed where required); Studio Sound run once per source; multicam/captions/layout dressed on-brand; the interview rule held (meaning + clip context preserved, approval offered, the human owning the final cut of anyone else's words); one corrected transcript shipped as the episode + human-picked clips + show notes + chapters, routed onward and published via WoopSocial; the workflow credit-aware under the post-Sept-2025 model (import only what you'll edit; verify tiers in-app); API/MCP automation only where actually connected; no cloned third-party voices, no meaning-flips, no fabricated tiers/credits/metrics; and correctly distinguished from capcut, captions-and-clipping/opus-clip, ai-voiceover, and podcast-and-audiograms.
以文字编辑的速度完成、并经试听验证的口播内容剪辑：先校正转录文本（姓名和术语——错误会造成错误剪辑和字幕），剪掉离题内容和失败录制，并通过批量 Underlord 处理移除填充词和静音；在文本中完成结构重组，并对结果进行完整试听；仅在词或短语的长度上，使用经同意验证的自有声音 Overdub 修正口误（整段内容应重录；在要求时披露合成语音）；每个源文件运行一次 Studio Sound；按品牌规范包装多机位画面、字幕和布局；遵守访谈规则（保留原意和片段上下文，为重大编辑提供审批选项，他人言论的最终剪辑决定权归人所有）；将一份校正后的转录文本作为单集节目 + 人工挑选的片段 + 节目说明 + 章节交付，继续分发并通过 WoopSocial 发布；在 2025 年 9 月后的计费模型下，让工作流关注 credits 消耗（只导入实际要编辑的素材；在应用内核实套餐）；仅在实际已连接时使用 API/MCP 自动化；不克隆第三方声音、不改变原意、不编造套餐、credits 或指标；并与 capcut、captions-and-clipping/opus-clip、ai-voiceover 和 podcast-and-audiograms 正确区分。