视频总结提示词库：12 条可直接复制的字幕稿提示词（2026）
想让 AI 把视频总结得好用，关键不是「帮我总结一下」，而是一条写死了输出结构、长度上限、以及「字幕稿里没提到就写未提及」的提示词。下面 12 条各管一种活，每条都在可一键复制的代码块里，拿走就能用。
12 条提示词正文一律保留英文——这是你真正要粘贴的内容，英文指令在各家 AI 对话工具里执行得更稳定；AI 回你的语言由你在对话里指定即可，不影响中文输出。
用法只有三步：先拿到视频的字幕稿，再复制其中一条提示词，最后把字幕稿替换掉末尾的 <paste the transcript here>。字幕稿从哪来？把视频链接丢给 BibiGPT 就能拿到带时间戳的文字稿，再粘到你常用的 AI 对话工具里跑这些提示词。
一条提示词值不值钱，不看措辞漂不漂亮，看它有没有替你规定输出长什么样。
带时间戳的字幕稿比纯文字稿好用得多——本文有一半提示词要求 AI 回引时间戳，没有时间戳它就只能写「n/a」，输出的可核查性会直接掉一半。如果你的字幕稿没有时间戳，这几条仍然能跑，只是少了跳转能力。
带时间戳的视频字幕稿与章节化阅读界面
视频很长，怎么快速找到该看的部分？
这件事由两条提示词负责：一条把逐字稿变成可跳转的目录，另一条把操作类内容压成编号步骤。视频很长、而接下来一小时不值得从头看到尾时，就先跑它们。两条成败都系于同一句约束：只用真实存在的时间戳，绝不自己编。
怎么让 AI 生成带时间戳的视频大纲？
先跑这一条，再决定要不要看、从哪里开始看。它输出 6 到 12 个带时间戳的小节，每节配一句说明，等于给这个视频做了一份跳转目录。关键约束是「只用字幕稿里真实出现的时间戳」——没有这句，AI 很容易顺手编一个看起来很合理的 00:14:32 出来。
You are given a video transcript with timestamps. Build a navigable outline of it.
Rules:
- Produce 6-12 sections. A section is one topic shift, not one paragraph.
- Format every line as: [HH:MM:SS] Section title - one sentence on what is covered.
- Use ONLY timestamps that appear in the transcript. Never estimate, interpolate or invent one. If a section starts where no timestamp is marked, write [time not marked] instead of guessing.
- Titles must be specific ("Why the 2019 pricing change backfired"), never generic ("Introduction", "Key points").
- End with one line: "Longest section: <title> (~<minutes> min)".
- Under 300 words. No preamble, no closing remarks.
TRANSCRIPT:
<paste the transcript here>怎么把教程视频变成一步一步的操作清单？
边看教程边来回拖进度条，是最费时间的看法。这条把教程压成编号步骤，一步一个动作，把口播提到的具体数值、命令和菜单路径原样抄进方括号。只在画面上出现、没被念出来的值会标成「shown on screen」而不是瞎猜，末尾还会补上前置条件、阶段性验收点和可能因版本而异的步骤。
Convert this tutorial transcript into a checklist someone can follow without watching.
Output numbered steps. Each step is one imperative action, followed in brackets by any exact value, setting, command or menu path that was spoken.
Rules:
- One action per step. Split any step that contains "and then".
- Copy values, names, paths and commands character for character. If something was only shown on screen and never spoken, write [shown on screen, not stated] - never guess it.
- After the steps add PREREQUISITES (what you need before step 1) and CHECKPOINTS (how you know each phase worked, quoting what the tutorial says you should see).
- Finish with one line: "Steps that may depend on the version you are using: ...".
TRANSCRIPT:
<paste the transcript here>
把视频字幕稿改写成文章大纲的操作界面
怎么把一个视频讲给不会打开它的人？
「讲给别人听」其实是两种活，取决于听的人是谁。要据此拍板的人需要一份结构固定的短结论；完全不熟这个领域的人需要术语被清掉、顺序被重排。下面两条各管一种，按对方的知识水平选，而不是按篇幅长短选。
不打算看完整段视频，怎么让 AI 写一份给决策者的摘要？
这条是写给「不会点开视频、但要拿这个做判断」的人看的。它固定四块：一句话结论、覆盖了什么、谁该看、以及视频没回答的问题。最后一块是多数摘要都缺的——它让读者知道自己还需要去别处补什么，而不是误以为看完摘要就掌握全貌了。
Summarize this transcript for a busy reader who will never watch the video.
Output exactly these four blocks:
1. BOTTOM LINE - 2 sentences. The single most useful thing in the video.
2. WHAT IT COVERS - 4-6 bullets, each under 20 words.
3. WHO SHOULD WATCH - 1 sentence on who benefits, 1 sentence on who can safely skip it.
4. WHAT IT DOES NOT ANSWER - 2-3 questions a reader may arrive with that the transcript leaves open.
Rules:
- Under 250 words in total.
- Every statement must be traceable to the transcript. If something is only suggested, write "implied, not stated".
- Add no background knowledge of your own.
TRANSCRIPT:
<paste the transcript here>完全不懂这个领域，怎么让 AI 讲成人话？
跨领域看视频，卡住的往往是几个术语。这条把内容重写成零术语版本，附一份人话词表，再按初学者需要的顺序（不一定是视频原本的顺序）重排 5 到 8 段，最后给一个明确标注为类比的比喻。讲者默认你已经会的东西，它会直说「视频假设你已经懂 X」，而不是自己编一段补上。
Rewrite this transcript for someone completely new to the subject.
Output:
- WHAT THIS IS ABOUT - 3 sentences, no specialist vocabulary at all.
- GLOSSARY - every specialist term the speaker uses, each explained in under 15 plain words.
- THE IDEA, STEP BY STEP - 5-8 short paragraphs, ordered the way a beginner needs them, which may differ from the order in the video.
- ONE ANALOGY - a single comparison to something ordinary, clearly labelled as an analogy.
Rules:
- Aim at a motivated 15-year-old reader.
- Where the speaker assumed background the transcript never supplies, write "the video assumes you already know X" rather than filling the gap yourself.
- No sentence longer than 25 words.
TRANSCRIPT:
<paste the transcript here>
由视频内容自动生成的自测题界面
怎么从一场会议里拿到决议和下一步？
开完会有两个各自独立的问题：要做什么，以及到底定下来了什么。第一条提示词给出带负责人的任务表；第二条划出「已决」和「只是讨论过」的界线，而录音恰恰会把这条线抹平。同一份逐字稿上都跑一遍，会后跟进就齐了。
怎么从会议录音里提取待办事项和负责人？
开完会或听完通话直接跑这条，拿到一张「事项 / 负责人 / 期限 / 时间戳 / 把握程度」的表。它明确要求没点名的事项写「unassigned」而不是顺着上下文猜人，并把「听起来像任务但其实被否掉了」的内容单列出来——后者才是会后扯皮的高发区。
Extract the action items from this transcript.
Return a table with the columns: Action | Owner | Due or trigger | Timestamp | Confidence
Rules:
- Owner: use the name actually spoken. If nobody was named, write "unassigned" - do not infer an owner from context.
- Due or trigger: the exact wording used, otherwise "not stated".
- Timestamp: where the action was raised. If the transcript carries no timestamps, write "n/a".
- Confidence: high (explicitly agreed) / medium (proposed, not confirmed) / low (mentioned in passing).
- Under the table add a NOT ACTIONS list: anything that sounds like a task but was rejected or postponed.
- If there are no action items, reply "No action items found" and stop. Do not manufacture any.
TRANSCRIPT:
<paste the transcript here>怎么分清会议里「已经定了的」和「还没定的」？
长会议的麻烦不是记不住，而是分不清「讨论过」和「定下来了」。这条把内容硬拆成已决、未决、仍在分歧三块，并且明确规定：只讨论没结论的一律算未决。中途被岔开、后面再没回来的话题也会被点名，这类漏项通常两周后才会被人发现。
Read this meeting transcript and separate what was settled from what was not.
Output three sections:
DECIDED - one line each: the decision, who stated it, the timestamp. Only include items where someone actually concluded something.
OPEN LOOPS - the question left hanging, who raised it, and what would close it.
DISAGREEMENTS - positions still in conflict when the topic changed. One sentence per side.
Rules:
- A topic that was merely discussed is an open loop, not a decision. Be strict about that line.
- If a topic was dropped and never revisited, say so explicitly.
- Maximum 8 items per section; if more exist, keep the most consequential and note how many you left out.
- Never assign a decision to a speaker who did not state it.
TRANSCRIPT:
<paste the transcript here>凡是没写「查不到就写未提及」的提示词，迟早会给你编一个时间戳。
怎么用一段课程录像来学习？
拿录像学习分成两件事：检验自己，以及留住打动你的部分。一条从素材里造出回忆型题目，另一条把值得逐字抄下来的句子挑出来。两条都不许用课外知识补空，所以结果始终忠实于这堂课本身。
怎么把网课视频变成一套自测题？
把一节课变成主动回忆的练习题，比重看一遍有效得多。这条会出 8 道回忆题、4 道需要串联两个知识点的应用题，外加带时间戳的答案。所有题目必须能只靠字幕稿作答，超纲的会被丢掉，所以它测的是这节课本身，不是模型的知识储备。
Turn this lecture transcript into a self-test.
Produce:
A. 8 recall questions (short answer) spread across the whole lecture, not only the opening third.
B. 4 application questions that require combining two ideas from the lecture.
C. An answer key. Each answer paraphrases the transcript and cites the timestamp it came from.
Rules:
- Every question must be answerable from the transcript alone. Delete any question that needs outside knowledge.
- Mark anything the speaker called important or exam-relevant with [flagged by speaker].
- Keep each answer under 40 words.
- If the transcript is too thin for 12 questions, produce fewer and say how many the material actually supports.
TRANSCRIPT:
<paste the transcript here>怎么从视频里挑出可以直接引用的金句？
做内容分发时最怕引错话。这条挑出 8 到 12 句可独立成立的原话，配时间戳和一句「为什么这句立得住」，并且死守逐字照抄：可以删语气词并用省略号标出，但不许润色。脱离上下文就变味的句子会被放进单独一栏，提醒你别拿去单独引用。
Pull the most quotable lines out of this transcript.
Return 8-12 quotes. For each: the quote verbatim, its timestamp, and one line on why it lands (surprising, contrarian, unusually concrete, memorable phrasing).
Rules:
- Verbatim means verbatim. You may cut filler words and mark the cut with "...", but never rewrite, smooth or improve the wording.
- Each quote must stand on its own without the surrounding context.
- Skip anything shorter than 6 words or longer than 40.
- If a line changes meaning once removed from its context, do not list it as a quote - put it under CONTEXT-DEPENDENT instead.
TRANSCRIPT:
<paste the transcript here>
从视频里标记并摘录金句的界面
怎么核对视频里到底说了什么？
引用或转发视频里的内容之前，值得先跑两遍：先把数字抓出来，再列出哪些说法需要核实。两条都不许判断真假，也不许生出没被说过的数字——目标是一份能逐条查的清单，而不是叠在原结论上的第二层结论。
怎么只提取视频里的数字和数据，还带时间戳？
财经、研报、行业分析类视频，你要的往往只有数字。这条把每个数字单独成行，附上原话、口径、时间戳和「有没有说来源」，并要求原样照抄——不四舍五入、不换算单位、不替讲者把「大概」这类模糊词去掉。没说来源的数字会被单独挑出来提醒你核实。
Extract every quantitative claim from this transcript.
One row each: Figure | Exact quote | What it measures | Timestamp | Source named? (yes: <source> / no)
Rules:
- Include numbers, percentages, dates, durations, amounts of money, counts and rankings.
- Quote the figure exactly as spoken. Do not round, convert units or normalize currencies.
- Keep any hedge the speaker used ("roughly", "about", "up to").
- Add no figure that is not spoken in the transcript, and do no arithmetic of your own.
- After the table, list under CHECK THESE any figure with no named source or that looks implausible, with one line saying why.
TRANSCRIPT:
<paste the transcript here>转发之前，怎么先列出视频里需要核实的说法？
注意这条不做核查，它只列出「哪些说法需要被核查」。每条输出说法、原话、时间戳、类型、怎么验证、以及说错了风险多大，按风险从高到低排。它被明确禁止用自己的知识判断真假——这样你拿到的是一份待办清单，而不是一层新的、同样需要被核查的结论。
Do not fact-check this transcript. Instead, build the list of things that WOULD need checking.
One row per checkable claim: Claim (one sentence) | Exact quote | Timestamp | Type (statistic / causal / attribution / prediction / definition) | How to verify it, in one line | Risk if wrong (high / medium / low)
Rules:
- Only falsifiable claims qualify. Tastes and preferences do not; predictions about measurable outcomes do.
- Do not say whether a claim is true or false, and do not use your own knowledge to judge it.
- If a claim is attributed to a study, person or report, put that attribution in the quote column exactly as spoken.
- Sort by risk, highest first. Maximum 15 rows.
TRANSCRIPT:
<paste the transcript here>怎么把一个视频变成另一种东西？
二次利用无非两种：把一个视频改写成别的形式，或者把两个视频摆在一起对照。第一条给出带论据的文章大纲；第二条把两份逐字稿的共识与分歧摊开。两条都拒绝补充原文里没有的证据。
怎么把一场演讲改写成文章大纲？
把一场演讲变成一篇文章，卡点在结构而不在文字。这条输出三个备选标题、一段导语、5 到 7 个「以观点而非话题命名」的小标题，每个小标题下配带时间戳的素材。最有用的是末尾的 GAPS：演讲没提供、但文章非补不可的数据和反方观点。
Turn this talk transcript into an outline for a written article.
Output:
- 3 candidate titles, each under 12 words.
- A lead paragraph of 60-80 words stating the article's argument.
- 5-7 section headings, each phrased as a claim rather than a topic label.
- Under each heading, 2-3 bullets of supporting material taken from the transcript, with timestamps.
- A GAPS list: what the article still needs that the talk does not supply (data, counterargument, example).
Rules:
- Invent no supporting evidence. Every bullet must exist in the transcript.
- Keep the speaker's argument, not your own opinion.
TRANSCRIPT:
<paste the transcript here>怎么让 AI 对比两个讲同一话题的视频？
同一个话题、两个人讲得不一样时，用这条把差异摆平。它分成共识、分歧、各自独有、以及「你该看哪个」四块，并强制标注是 A 说的还是 B 说的。它还允许 AI 承认「只是措辞不同、看不出实质分歧」——这一句避免了大量被硬造出来的对立。
You are given two transcripts on the same topic, labelled A and B.
Output:
1. AGREE - points both make, one line each, with a timestamp from each side.
2. DISAGREE - A's position, B's position, and the evidence each one offers.
3. ONLY IN A / ONLY IN B - what one covers that the other never raises.
4. WHICH TO WATCH - 3 sentences: for which reader, which video, and why.
Rules:
- Never merge the two into a single voice. Always attribute.
- A point that appears in only one transcript does not belong in AGREE.
- Where you cannot tell whether they truly disagree or merely word things differently, write "wording differs, substance unclear".
TRANSCRIPT A:
<paste transcript A here>
TRANSCRIPT B:
<paste transcript B here>
把字幕稿粘进 AI 对话窗口继续追问
怎么把这些提示词改成自己的？
改三个地方就够：长度上限（250 字太短就往上调）、输出块的名字（换成你团队在用的说法）、以及表格列（加一列「关联项目」之类）。唯一不建议动的是每条最后那几行约束——「只用真实出现的时间戳」「没提到就写 not stated」「不许自己补背景知识」，这三句才是让输出可以直接拿去用的原因。
12 条里你真正每周都会用的，通常不超过 3 条——先把那 3 条固定下来。
把常用的几条存进输入法短语或者代码片段工具，配上两三个字母的缩写，比收藏这篇文章更能保证你真的会用。用顺手之后再回来挑第 4 条，比一次性全套上更容易坚持。