Version:v1.1 Use Cases: Check the paragraph transitions, information density, and spoken delivery fluency of a short video verbatim script, and identify the risk points where "viewers will swipe away at which second"
Output Boundaries
Use plain Markdown to deliver the review report and revision record in the current conversation, keeping it platform-independent. Do not rely on a specific note-taking app, and do not create, save, or open local review files; file rendering and application opening are handled by capabilities outside the project.
Trigger Conditions
When the user:
- Says "check logical continuity", "see if the logic is broken", "is there anything verbose"
- Says "help me see if this script flows smoothly", "can this script keep people watching"
- Says "check completion rate risk", "where will viewers swipe away"
- Uses
/dbs-script-flow
Differences from Other Skills
| Skill | What to Check | Analogy |
|---|---|---|
dbs-script-flow Other Skills besides |
Each is invoked by its own English standard name | Check according to the corresponding task |
/dbs-script-flow |
Micro transitions: whether each paragraph flows smoothly into the next | Check whether any step in the staircase is broken |
Core Principles
The enemy of completion rate is not bad content, but a moment when viewers can no longer follow.
Viewers swipe away for only three reasons:
- Logical break: The previous paragraph talks about A, and the next paragraph suddenly jumps to C, with B missing in between
- Information density dropped: A certain paragraph repeats what has already been said, or goes in circles
- Spoken delivery stumbles: Sentences are too long, too written, or self-questioning and self-answering, making them impossible to read aloud
Review Process
Step 1: Segment
Divide the verbatim script into several segments according to natural paragraphs (topic switch points), and label each segment with:
- Segment number
- Segment topic (summarize in one sentence what this segment is talking about)
- Segment function (opening / background setup / argument / case / transition / climax / ending)
Step 2: Scan each segment across three dimensions
For each segment, check in order:
Dimension 1: Logical connection (between adjacent segments)
Ask yourself:
- Is there a logical jump between the last sentence of the previous segment and the first sentence of this segment?
- After hearing the previous segment, can viewers naturally understand why this segment is being said?
- If not, what is the missing piece of information that bridges the two?
Common problems:
- Suddenly jumping from an individual case to macro analysis, with a transition sentence missing in between
- Suddenly jumping from explaining a cause to another topic, without closing the previous topic
- Jumping from data to opinion, with a "therefore" missing in between
Fix methods:
- Add a transition sentence (usually one sentence is enough)
- Add a closing sentence at the end of the previous segment
- Adjust the order of segments
Dimension 2: Information density (within the segment)
Ask yourself:
- Does this segment repeat information already stated earlier?
- Is this passage beating around the bush (using 5 sentences to say something that could be said in 1)?
- If this passage were deleted, would the audience miss key information?
Common problems:
- The same meaning is repeated in three different ways
- When explaining a concept, too many examples are given—the audience already understands but examples keep coming
- A certain passage is "correct but empty"—saying it is the same as not saying it
Fixes:
- Cut repeated phrasing and keep only the most powerful version
- Merge redundant passages
- Mark as "can be deleted" and let the user decide
Dimension 3: Spoken delivery fluency (sentence level)
Ask yourself:
- Can this sentence be read out in one breath? (Be wary of single sentences over 30 characters)
- Is there written-language phrasing? ("aimed at", "in view of", "based on the above analysis")
- Is there self-questioning and self-answering? ("Do you think X? Actually Y"—occasionally it's fine, but using it continuously gets annoying)
- Are there invented concepts? (Terms the audience can't understand, with no immediate explanation)
Common problems:
- Modifiers stacked on modifiers, cramming too much information into one sentence
- A term is used but not explained right away
- Several short sentences in a row make the rhythm too choppy, or several long sentences in a row make the rhythm too dull
Fixes:
- Split long sentences into two
- Follow a term immediately with a plain-language explanation
- Adjust the rhythm of alternating long and short sentences
Step 3: Mark risk levels
Mark a risk level for each problem found:
| Level | Meaning | Description |
|---|---|---|
| 🔴 High risk | The audience will most likely swipe away here | Logical break, severe drop in information density |
| 🟡 Medium risk | The audience may zone out but won't necessarily swipe away | Transition not smooth enough, slightly verbose |
| 🟢 Minor suggestion | Better if changed, fine if not | Minor spoken-delivery tweaks, wording optimization |
Output format
## 逻辑延续检查报告
### 段落结构
| 段号 | 主题 | 功能 | 状态 |
|-----|------|------|------|
| 1 | [主题] | 开场 | ✅ |
| 2 | [主题] | 背景铺垫 | ✅ |
| 3 | [主题] | 论证 | 🟡 |
| ... | ... | ... | ... |
### 问题清单
**🔴 问题 1:第 X 段 → 第 Y 段,逻辑断裂**
> 原文(第 X 段末尾):「……」
> 原文(第 Y 段开头):「……」
**问题**:[具体描述为什么这里会让观众跟不上]
**修复建议**:在第 X 段末尾加一句过渡:
> 「[具体的过渡句]」
---
**🟡 问题 2:第 Z 段,信息密度下降**
> 原文:「……」
**问题**:[具体描述为什么这里观众会走神]
**修复建议**:
> 改为:「[精简后的文本]」
---
**🟢 问题 3:第 W 段第 N 句,口播不顺**
> 原文:「……」
**问题**:[具体描述为什么这句话念不出来]
**修复建议**:
> 改为:「[改写后的句子]」
---
### 整体评估
- **逻辑衔接**:[一句话总结]
- **信息密度**:[一句话总结]
- **口播流畅度**:[一句话总结]
- **总结**:[整体判断,需要改几处,改完可以发]
### 按已有请求决定是否改稿
用户在当前请求或前文已经说「帮我改」「直接改」等,直接进入标记式改稿,不再次询问授权。入口编排转交的用户原始请求同样有效。
用户只要求检查、尚未表达改稿意图时,在整体评估后可以询问:
> 要直接帮你标记式改稿吗?改的时候保留你的原文,用删除线和 🆕 emoji 标出改动。
用户明确只要检查、不希望改稿或讨论下一步时,交付检查结果即停止。没有改稿请求时,不自行改写原文。
Markup-style revision workflow (executed after user confirmation)
When the user says "revise," "help me revise," "just revise it directly," or "revise the original text" in the current message or earlier context, start markup-style revision without replying for confirmation again.
Core principles
Never erase the user's original text. Revision means overlaying marks on the original text, not replacing the old version with a new one. The user must be able to see at a glance:
- Which parts they wrote themselves
- Which parts are AI deletion/replacement suggestions
- Which parts are content added by AI
- Which parts were previously changed by AI and then restored (to avoid losing history across multiple rounds of feedback)
Three general Markdown markers
| Symbol | Purpose | Example |
|---|---|---|
~~删除线~~ |
Mark original text suggested for deletion or replacement | ~~重点是在于,这里划重点啊~~ |
| 🆕 | Mark newly added/rewritten content | 🆕 这里有个关键点 |
| ⚠️ | Mark where the AI retracts its own judgment and restores the original text | ⚠️ 上面的删除判断错了,已还原。这三段不是跑题,是在拆解…… |
Revision rules
- Make changes directly at the original positions—do not write a complete new version and paste it at the bottom; mark in place
- Every change appears in pairs——
~~原文~~Follow with 🆕 new version, so the comparison is clear at a glance - When deleting an entire passage: keep the original passage with strikethrough, and below use 🆕 to explain why it was deleted (or give a condensed replacement)
- After finishing, give a change list at the end— in table form, listing the location, type, and reason for each change, so the user can quickly review
- Multiple rounds of feedback should append, not overwrite— if the user points out that something was changed incorrectly, mark the restoration with ⚠️; do not just change it back as if nothing happened
Output format after revision
[原文 + 标记式改动,原地操作]
---
### 改动清单
| 位置 | 类型 | 说明 |
|---|---|---|
| 开头 | 🟢 替换 | [简短说明] |
| 第 X 段 | 🔴 整块删除 | [原因] |
| 结尾 | 🟡 新增 | [新增了什么、为什么] |
At the end, give a one-sentence verdict (which parts must be changed, which parts are optional), so the user knows the priorities.
Notes
- Do not change the content or viewpoints: this skill only handles "how to say it," not "what to say." Do not change the user's arguments, cases, or data
- Fix suggestions should be conversational: the transition sentences and rewrites provided must be directly speakable, not written-style
- Do not over-diagnose: if the draft is overall very smooth, just say "no problems," and do not force yourself to find faults
- Prioritize high-risk issues: better to miss a minor suggestion than to drown high-risk issues in a pile of small ones
- Keep the user's tone: when rewriting, keep the user's own speaking style; do not turn the user's words into an "AI flavor"
- Give specific locations: every issue must quote the original text, so the user can find it at a glance
- Confirm only when the shooting method affects a specific revision: first use the available information. If the proposed change involves parallel lists, board writing, gestures, or visual coordination such as "highlight this" or "look here," and different shooting methods would change the deletion/revision conclusion, ask clearly about the shooting method before revising these locations. Lists and deictic words in whiteboard videos may serve a necessary function and should be kept. Logical jumps, redundancy, and long sentences unrelated to the shooting method can be handled first; do not pause revision of the entire script merely because the shooting method was not specified, and do not assume on your own that it is pure voiceover.
- Before judging something as "off-topic," ask one question: if this passage is deleted, can the audience infer the specific meaning on their own from the previous key point? If they cannot infer it, then this passage is necessary argumentative substance, not excess fat. Do not only look at "whether the topic is literally a bit off"; look at "which argument this passage serves"
Practical cases
Case: "Xiaohongshu Was Sued" verbatim script (2026-04-09)
🟡 Issue 1: Paragraph 5, line 46, the spoken delivery is convoluted
Original: "You will then enter another interval similar to what is called a merchant traffic pool"
Issue: "another interval similar to what is called..." has two consecutive "a" particles, so it catches when read aloud. Also, "merchant traffic pool" is a technical term, and the audience may not understand it.
Fix suggestion:
Change to: "You just won't have much traffic. It's not that there's none at all, but it will be very poor." State the result directly, without introducing a concept.
🟡 Issue 2: Paragraph 7 → Paragraph 8, missing transition
Original (end of paragraph 7): "So next is that you violated the rules by directing traffic, and the platform bans your account." Original (start of paragraph 8): "And then what happened next was even more interesting"
Issue: Jumping directly from "lawyer banned" to "courtroom speech" lacks a transition in between. The audience is still digesting the information that the "lawyer was banned" and is suddenly pulled into the courtroom scene.
Fix suggestion: Change the end of paragraph 7 to:
"So what this lawyer encountered next, being throttled, is in fact the standard outcome of this process." "Standard outcome" wraps up the preceding analysis, then opens up new information.
End directly after completing the current task. Only when the user explicitly asks about the next step and the current environment already has /dbs installed, briefly prompt: "When unsure about the next step, you can enter /dbs。」