Diagnose first, then rewrite. The goal is not to mechanically insert questions, numbers, or exaggerated promises, but to make the first beat provide a reason to watch and the second beat continue providing new information.
The empirical baseline of this skill comes from the current account's 72 valid Douyin videos, 40 linkable verbatim transcripts, and 38 records of shooting format. It is suitable for guiding this account's AI, tool, and knowledge-style spoken videos, and does not package empirical correlation as universal causation across accounts.
Before each use, read references/rubric.md, and execute according to its scoring boundaries, counterexamples, and data definitions.
Applicable Inputs
Accept any of the following inputs and proceed directly, without blocking due to incomplete information:
- A complete verbatim transcript or subtitles with time codes;
- Copy for 0–2 seconds and 2–5 seconds explicitly marked by the user;
- A complete spoken script without time codes;
- Only the first few sentences, accompanied by the title, audience, or platform;
- The user only requests generating a new opening and does not request showing a detailed analysis first.
If the user provides a title, check whether the spoken content serves as a "sequel to the title" rather than repeating the title. If there is no title, skip this item and note that title continuation cannot be judged.
Step 1: Read through the body and establish the fact boundary
First read through the complete verbatim transcript and extract content that actually exists in the body:
- The conclusion the video ultimately delivers;
- Verifiable numbers, test scope, and social proof;
- The specific result viewers can obtain;
- The first step, demonstration, case, or before-and-after comparison;
- The creator's real experience and expression style.
Rewriting may only use these facts. Do not fabricate test participant counts, effects, prices, brands, rankings, authoritative endorsements, or user feedback in order to be "more impactful." When the body lacks evidence, use specific but non-quantified benefits, or label [需补真实证据]。
Step 2: Confirm the first-5-seconds boundary
Handle according to the following confidence levels:
- High confidence: There are time codes, or the user explicitly marks the actual 0–2 seconds and 2–5 seconds.
- Medium confidence: There is video, audio, or an explicit statement of which sentence ends at the 5th second.
- Low confidence: Only a verbatim transcript without a timeline. Use the first effective sentence and the first two natural spoken beats as structural proxies, and clearly state that this is "not an actual duration measurement."
Do not hard-cut at 2 seconds and 5 seconds based on a fixed number of Chinese characters, average speaking rate, or punctuation count. Everyone's speaking rate, pauses, and visual coordination differ.
If key information comes from the first frame, subtitles, facial expressions, on-screen demonstrations, or before-and-after comparisons, a pure verbatim transcript can only evaluate the "spoken opening." List the missing visual information, but still complete a provisional analysis.
Step 3: Diagnose the two beats separately
0–2 seconds: grab power
Judge whether viewers can immediately know:
- Who or what tool or what scenario this content is about;
- What specific problem, loss, conflict, twist, or choice is happening;
- What new information the first sentence provides compared with the title;
- Even without conflict, does it directly reveal a specific result worth watching.
The task of the first beat is to give a reason to watch, not to introduce this episode's agenda.
2–5 seconds: Retention power
Determine whether the second beat continues to advance:
- Provides new evidence, an answer, a concise mechanism, or the first segment of a demonstration;
- Gives a clear, credible, verifiable benefit;
- Does not repeat the title or the first beat;
- Does not suddenly start reading a long prompt, course outline, term definitions, or long list;
- Can naturally connect to the first step or main content after 5 seconds.
A low 2-second skip rate does not equal a high 5-second completion rate. When the first beat is very strong but the second beat falls into a long prompt or outline, it must be clearly judged as "strong at 2 seconds, weak retention at 5 seconds".
Step 4: Identify the biggest single-point problem
From the following problems, pick only the 1–3 that most affect retention, sorted by priority:
- Title repetition with no new information;
- Agenda-style meta-talk: "Today I'll share" "This video will make it clear" "First, understand";
- Definition first, user benefit appears too late;
- Problem too broad, with no specific consequence;
- Has numbers or "personally tested", but no conclusion, standard, or benefit;
- The second beat repeats the first beat with no semantic advancement;
- The cognitive load of the second beat spikes sharply;
- Promise exceeds what the main content can deliver;
- The opening target does not match the audience of the main content;
- Visual evidence may be important, but is missing from the current input.
Do not treat form as conclusion:
- A question itself adds no points;
- Numbers themselves add no points;
- "I tested it" or "I've done this for ten years" adds no points by itself;
- Naming tools such as Claude, API, or GitHub does not by itself make the scenario specific;
- "Make it clear" appearing in the title does not have to be mechanically marked wrong; the real problem is whether the spoken script continues to repeat it without adding new facts.
Step 5: Rewrite
By default, provide two versions:
- Minimal-change version: Preserve the original tone, facts, and sentence structure as much as possible, only moving truly valuable information forward.
- Enhanced reconstruction version: Explicitly write it as "0–2 seconds / 2–5 seconds", so the two beats take on different tasks.
If the original opening is already very strong, do not force a rewrite just to pursue flashiness. Point out what should be kept, and only make necessary cuts or optimize the second beat's retention.
Only when the user explicitly requests "5 options", "more options", or multi-platform alternatives, provide 5 differentiated versions. The differences should come from pain points, twists, leading with results, action instructions, real evidence, or scenario entry points, rather than swapping synonyms in the same sentence.
After rewriting, self-check item by item:
- Whether it preserves facts the main content can deliver;
- Whether the first beat adds new information;
- Whether the second beat continues to advance;
- Whether the agenda, credentials setup, and long outline are removed;
- Whether questions, numbers, or exaggerated words were mechanically inserted;
- Whether it can be said naturally and flow into the main content.
Step 6: Provide shooting coordination
Only give shooting suggestions directly related to the first 5 seconds:
- People, conflicts, keywords, or results suitable for presentation in 0–2 seconds;
- At 2–5 seconds, you can cut to real data, interface, error messages, before-and-after comparisons, or the first step of an operation.
"The person appears on camera first, then cut to evidence or demonstration at 2–5 seconds" is a hypothesis worth verifying in the current data, not a hard rule. There is no reliable difference between landscape and portrait orientation; do not add or subtract points from opening copy based on screen orientation.
Output format
Output according to the following structure by default. When information is limited, the length can be compressed, but retain the section-by-section assessment and at least one version that can be used directly.
# 前 5 秒诊断
- **总结判定**：合理 / 基本合理 / 高风险 / 需要重构
- **边界置信度**：高 / 中 / 低；说明本次如何划分两拍
- **评分**：0–2 秒 X/40；2–5 秒 X/40；表达与兑现 X/20；总分 X/100
- **一句话主因**：最影响抓取或承接的问题
## 分段诊断
| 阶段 | 原文 | 当前功能 | 有效点 | 问题 | 观众此刻还缺什么 |
|---|---|---|---|---|---|
| 0–2 秒 | … | … | … | … | … |
| 2–5 秒 | … | … | … | … | … |
## 优先修改
1. [影响 2 秒抓取 / 5 秒承接] …
2. …
## 改写
### 最小改动版
- **0–2 秒**：…
- **2–5 秒**：…
### 强化重构版
- **0–2 秒**：…
- **2–5 秒**：…
### 为什么这样改
- …
## 拍摄配合
- **0–2 秒画面**：…
- **2–5 秒画面**：…
## 发布后验证
- 本账号参考线：2 秒跳出率 32.69%，5 秒完播率 40.92%。
- 一次尽量只测试一个变量；参考线仅用于本账号历史对比，不是普遍行业标准。
Expression requirements
- Conclusions should be direct; do not use vague boilerplate such as "it may still be possible to further optimize."
- Clearly state whether the problem occurs at 0–2 seconds or 2–5 seconds.
- Explain what problem each modification addresses.
- Use precise wording such as "more likely to affect" and "there is a risk," and do not disguise statements as certain causation or predictions of specific view counts.
- Rewrites should be natural and directly speakable, not written as a pile of advertising slogans.