---
name: transcript-polish
description: "This skill should be used when the user wants to polish a raw spoken transcript from a recording into a clean verbatim script (逐字稿). It handles very long texts of 30000 or more characters by processing in chunks. Before starting, it asks the user 4 questions to calibrate: (1) polish intensity A/B/C, (2) style preference — oral/written/custom-described, (3) paragraph handling, (4) whether to generate a PPT outline. The PPT outline uses a 3-level structure: chapter titles, per-slide bullet points, and speaker notes. Output can be merged or separate. Trigger phrases: 润色口水稿, 整理逐字稿, 把录音稿整理成逐字稿, polish transcript, 整理录音稿, 润色讲稿."
---

# Transcript Polish — 录音口水稿润色为逐字稿（含 PPT 框架）

## Purpose

Polish raw spoken transcripts (口水稿) from recordings into clean, publishable verbatim scripts (逐字稿), and optionally generate a structured PPT outline from the same content.

**Core principle**: Preserve original content length and information density. This is a *polish*, not a rewrite or compression. Style parameters are determined by the user before work begins.

---

## When This Skill Is Used

- User provides a raw spoken transcript from a recording
- User asks to "润色口水稿"、"整理逐字稿"、"把录音稿整理成逐字稿"
- Text contains heavy filler words, repetitions, and spoken-language artifacts

---

## Step 0: Ask the User 4 Questions Before Starting

**ALWAYS do this first**, before reading or processing any transcript content.

Present the following four questions together in one message, clearly formatted:

---

> 在开始之前，我需要了解你的偏好，请回答以下 4 个问题（直接用序号回答即可）：
>
> **① 润色力度**
> - A｜轻度 — 只去掉明显的填充词（嗯、啊、那个……），保留口语句式，读起来仍像讲话
> - B｜标准 — 去掉填充词 + 整理假起步和重复，句子流畅但不改变语气（推荐）
> - C｜深度 — 在标准基础上进一步规范句式，使其更接近书面表达，但不压缩内容
>
> **② 风格保留**
> - A｜保留口语风 — 保留"大家注意一下"、"就是说"、"对吧"等教学腔和说话节奏
> - B｜适度书面化 — 保留内容和逻辑，但去掉过多的口语化表达，更像文章
> - C｜我来描述 — 我会告诉你我想要的风格，按我说的来
>
> **③ 段落处理**
> - A｜保持原结构 — 不新增或合并段落，原文怎么分就怎么分
> - B｜优化段落 — 在话题切换处适当添加段落分隔，让层次更清晰
>
> **④ 是否同时生成 PPT 框架**
> - A｜是，逐字稿 + PPT 框架都要（同一份文档里，PPT 框架附在末尾）
> - B｜是，但分开输出（先给逐字稿，再单独给 PPT 框架）
> - C｜不需要，只要逐字稿
>
> 如果有任何特殊要求（比如保留某些词、PPT 面向的受众等），也可以一并告诉我。

---

After the user answers, **confirm the settings** in one line before proceeding:

> 好的，我将按照：**[力度]** × **[风格]** × **[段落]** 来处理，并**[生成/不生成] PPT 框架**，现在开始。

If the user skips the questions or says "随便" / "默认"，use: **标准 × 保留口语风 × 优化段落 × 生成 PPT（合并输出）**，and note this to the user.

---

## Polish Parameters Reference

### Dimension 1: Polish Intensity

| Level | What Changes | What Stays |
|-------|-------------|------------|
| **A 轻度** | 纯填充词（嗯/啊/呃/那个）| 所有口语句式、重复强调、假起步 |
| **B 标准** | 填充词 + 假起步 + 明显重复 | 讲师语气、口语表达、中英混用 |
| **C 深度** | 标准 + 口语句式规范化 | 所有内容、例子、信息量 |

### Dimension 2: Style Preference

| Option | Description |
|--------|-------------|
| **A 口语风** | 保留教学腔、说话节奏、"大家"等称呼，读起来像讲课 |
| **B 书面风** | 减少语气词和口头习惯，读起来像文章或讲义 |
| **C 用户描述** | 严格按用户提供的风格说明执行；若描述模糊，在开始前追问确认 |

#### 当用户选择风格 C 时的追问流程

如果用户选择了"C｜我来描述"，在确认其他参数后，**单独追问**：

> 好的，请描述你希望的风格。可以参考以下几个维度来描述（不需要都答，说你在意的就行）：
> - 目标读者是谁？（学生、同行、大众……）
> - 读起来应该像什么？（讲课、文章、对话、报告……）
> - 有什么特别要保留或避免的表达方式？
> - 有没有可以参考的风格样本（一段文字示例）？

收到用户描述后，用一句话复述你的理解并确认，例如：
> 明白了，我会整理成适合大众阅读的对话体风格，保留讲师的幽默感和举例方式，去掉过于学术化的表达。这样理解对吗？

### Dimension 3: Paragraph Handling

| Option | Description |
|--------|-------------|
| **A 保持原结构** | 不新增、不合并段落；原有换行照留 |
| **B 优化段落** | 在话题切换、例子开始/结束处补充段落分隔；长段可适当拆分 |

---

## What to Always Preserve (Regardless of Parameters)

- **All content**: Every example, story, case study, and technical point must remain
- **Original length**: Output character count must stay within ±15% of input
- **Logical sequence**: Do NOT reorder topics, points, or transitions
- **Technical terms and jargon**: Preserve exactly as spoken
- **No added content**: Do NOT introduce explanations, summaries, or commentary

---

## What to Always Clean (Regardless of Parameters)

These are removed at all intensity levels:

| Category | Examples |
|----------|----------|
| Pure filler sounds | 嗯、呃、啊、哦（standalone, no semantic value） |
| Transcription artifacts | [inaudible]、(停顿)、[PPT切换] |
| Identical word stutters | "这个这个这个" → "这个" |

---

## Handling Long Texts (30,000+ Characters)

Long transcripts MUST be processed in chunks to avoid truncation and quality degradation.

### Chunking Strategy

1. **Determine chunk boundaries**: Split on natural paragraph or topic boundaries (not mid-sentence). Target ~3,000–4,000 characters per chunk
2. **Apply user parameters consistently** across all chunks — note the user's style settings before chunk 1 and apply them throughout
3. **Track cross-chunk consistency**: Note recurring terms and the speaker's specific phrases from early chunks
4. **Reassemble**: Concatenate all polished chunks in order without adding or removing section breaks

### Workflow for Long Texts

```
Step 0: Confirm user's 4 parameters (see above)

Step 1: Scan the first ~1,000 chars to identify:
         - Speaker's domain and tone register
         - Recurring filler words specific to this speaker
         - Main topic areas (used later for PPT outline if needed)
         - Any special terms or patterns to handle consistently
         - If PPT is requested: start a running outline log (topic → key points)
           to accumulate chapter/slide candidates across chunks

Step 2: Divide into chunks at natural paragraph breaks

Step 3: For each chunk, apply parameters:
         - Remove fillers per intensity level
         - Adjust style per style preference
         - Handle paragraph breaks per paragraph setting
         - Verify no content was added or removed
         - If PPT is requested: after each chunk, append to the running outline log:
           note any new topic transitions, key concepts, and examples encountered

Step 4: Reassemble polished transcript and do a final pass for:
         - Consistent terminology throughout
         - No orphaned sentence fragments at chunk boundaries
         - Paragraph structure matches chosen setting

Step 5: If PPT framework is requested, generate it after
        the full transcript is assembled, using the running outline log
        accumulated in Steps 1–3 as a structural guide.
        The final PPT must still be derived from the polished transcript,
        not just from the notes — the log is a scaffold, not the source.
```

### Reporting Progress

When processing a long text, report progress after each chunk:
> "已完成第 N 段（约 XXXX 字），继续处理第 N+1 段…"

---

## PPT Framework Generation

Generate the PPT outline **after** the full polished transcript is complete, based on the polished text (not the raw input).

### Structure: Three Levels

```
# [演讲/课程总标题]

---

## 第一章：[章节名]

### Slide 1：[页面标题]
- 要点一
- 要点二
- 要点三
> 备注：[讲师在逐字稿中对应的讲解提示，1-2句，帮助演讲者记忆讲解要点]

### Slide 2：[页面标题]
- 要点一
- 要点二
> 备注：[讲解提示]

---

## 第二章：[章节名]
...
```

### PPT Generation Rules

1. **章节（Chapter）划分**：根据逐字稿中的主题切换确定，通常对应一个完整的知识模块或论点
2. **每页（Slide）划分**：对应逐字稿中一个具体的知识点、例子或论证步骤；每页要点 3-5 条为宜
3. **要点提炼**：从逐字稿中提取关键信息，用简洁名词短语表达（非完整句）；不添加原文没有的内容
4. **备注撰写**：从逐字稿对应段落中提取 1-2 句核心讲解句，帮助演讲者现场展开讲解；语气与逐字稿风格保持一致
5. **页数估算**：以每 800-1,200 字对应 1 页为参考（实际以话题完整性为准）
6. **封面页**：始终生成一个封面 Slide（标题页），包含课程/演讲标题和可选的副标题
7. **总结页**：如果讲师在原文中有明确的总结段落，生成一个总结 Slide

### Output Format by User Choice

**选项 A（合并输出）**：

```
========================================
【逐字稿】
========================================

[完整逐字稿内容]


========================================
【PPT 框架】
========================================

[PPT Markdown 大纲]
```

**选项 B（分开输出）**：
- 先完整输出逐字稿，末尾注明："PPT 框架已准备好，请回复「继续」获取。"
- 用户确认后，单独输出 PPT 框架

---

## Output Format

- Output as Markdown
- If **段落 B** is selected: use blank lines between paragraphs for clear separation
- Do NOT add headings in the transcript unless the speaker explicitly announced section titles
- Preserve any numbered lists or explicit structure the speaker used

---

## Reference Files

- See `references/filler_words.md` for a comprehensive list of common Mandarin filler words and patterns, organized by intensity level
- See `references/polish_examples.md` for before/after examples at each intensity level, and a PPT outline example

---

## Quality Check (Before Delivering Output)

**Transcript check:**
1. [ ] Character count of output is within ±15% of input
2. [ ] No new content was introduced
3. [ ] Polish intensity matches the user's chosen level
4. [ ] Style matches the user's preference
5. [ ] Paragraph structure matches the user's setting
6. [ ] All examples and case studies are present
7. [ ] No sentences are grammatically broken or truncated

**PPT framework check (if generated):**
8. [ ] Every slide's bullet points are drawn from the transcript — no invented content
9. [ ] Speaker notes are extracted from the transcript, not paraphrased from imagination
10. [ ] Chapter divisions match the actual topic structure of the transcript
11. [ ] Slide count is proportional to content length