ppt-narrative-review

ppt-narrative-review — Storyline Review

When to Trigger

Applicable：

• The argumentation has already been completed (ran ppt-classify + the corresponding argumentation skill)
• There is a drafted storyline / chapter outline / slide outline
• Before delivering visual output (such as visual-deck), you want to review the structure first

Not Applicable：

• The argumentation has not been done yet → run ppt-research-setup or plan-ceo-review
• A visual version of the PPT has already been produced and you want to review the visuals → use design-review(gstack) or visual-deck 's built-in page review
• You want the skill to write the storyline for you directly → this is a review skill, it does not generate

Core Principle: Type Determines the Reasonable Structure

A "good structure" for a pitch-type PPT becomes a disaster when placed in a research-type one. And vice versa.

Before reviewing a storyline, you must first know the PPT type. Different types have different "fitting structures":

Research-driven storyline (push → the audience walks along with you)

开场：反共识悖论
    ↓
框架：结构化坐标系
    ↓
调查路径 1（直观层）→ 产出阶段性证据
    ↓
调查路径 2（反推层）→ 产出反直觉发现
    ↓
调查路径 3（盲区层）→ 产出长期威胁/机会
    ↓
综合：多路径交汇的判断（不是开场就说的）

Characteristics:

• Conclusion comes later, the opening is a question, not a conclusion
• Layered evidence(intuitive / inferential / blind spots)
• Falsifiabilitywritten on its face ("if X happens, this judgment fails")

Pitch-driven storyline (pull → the audience is persuaded by you)

开场：结论前置（判断 + 赌注）
    ↓
痛点：为什么现在必须做
    ↓
论据 1 → 论据 2 → 论据 3（逐层加强）
    ↓
反方论据 + 反驳
    ↓
行动号召（下一步是什么）

Characteristics:

• Conclusion up front, tell the audience right at the start what you want them to accept
• Progressive arguments(not a parallel list)
• A clear Ask(want money / want a decision / want resources)

Teaching-driven storyline (teach → the audience learns how to do it)

开场：痛点（没有这个能力你会遇到什么问题）
    ↓
承诺：学完你会做到什么（能力目标）
    ↓
步骤 1 → 2 → 3（操作序列，可复现）
    ↓
常见陷阱 + 排错
    ↓
进阶 / 变体（如果还有时间）

Characteristics:

• Clear capability goal(what they can do after listening, not what they understand)
• Reproducible steps(the audience can complete them independently)
• Pitfalls up front(pitfall reminders go inside the steps, not at the end)

Narrative-driven storyline (carry → the audience accompanies you for a stretch of the journey)

开场：主角 + 日常状态
    ↓
激励事件（触发冲突的那一刻）
    ↓
挣扎 / 尝试 / 失败
    ↓
转折（顿悟 / 外部变化 / 代价）
    ↓
新状态（主角变了什么）
    ↓
对观众的意义（但不直接说，让观众带走）

Characteristics:

• Clear protagonist(not "we", but a specific person or system)
• A real conflict(not "the challenge is great")
• The turning point has a cost(not "and then we did it")

Review Dimensions

Dimension 1: Structural Fit

Compare against the "fitting structure" corresponding to the type, and score:

Status	Score	Handling
Fully matches	9-10	Proceed to the next dimension
Main structure is right, details deviate	6-8	Mark the deviation points, give specific suggestions
Uses the wrong type of structure (e.g., research uses a pitch opening)	1-5	Stop, tell the user "either change the structure or reclassify" (go back to ppt-classify）

Dimension 2: Introduction, Development, Turn, and Conclusion

Does every stage have something carrying it? Or does the middle section collapse?

Common collapse points：

• Opening collapse: the paradox / pain point / protagonist has no tension → the audience zones out in the first 30 seconds
• Middle collapse: among the three arguments or three paths, one is especially weak and drags down the whole
• Ending collapse: Everything before is buildup, and the ending is glossed over with a single line like "so the conclusion is xxx"

Dimension 3: Key information load-bearing

Are the core evidence / core judgment / core emotion presentVisual anchor points？

Visual anchor points = the few pages the audience remembers most:

• Opening page 1
• The transition page of each section
• The last 1-2 pages of the ending
• The pages with a key visual

Example problem: The hardest numbers are buried in a table on page 7 and never make it onto a key visual → suggest moving them to a transition page.

Dimension 4: Fact-checking (new in v0.2.0 · mandatory)

During the argument-building stage, AI will "confidently fabricate" technical terms / data / product models. This is not an occasional bug; it is a structural risk. For a research-type PPT, one fabricated number will be spotted instantly by a technical audience, and all credibility collapses.

During review, these 4 principles must be checked page by page:

4.1 · Codename + year, not "flagship"

❌ Not allowed:

• "NVIDIA's latest flagship GPU"
• "the previous-generation architecture"
• "the current main chip"

✅ Required:

• "NVIDIA H200 · Hopper · 2024"
• "Blackwell B200 · 2025"
• "Vera Rubin · 2H 2026"

Why：旗舰 is a tense-drift word—today's flagship is not tomorrow's. Codename + year never goes out of date.

4.2 · Percentages must have a clear numerator and denominator

❌ Not allowed:

• +73% YoY · 数据中心营收(the numerator and denominator do not match; the actual 73% is total revenue growth)
• 80% 份额(whose share? which market?)

✅ Required:

• $681B · +73% 总营收同比(put the absolute value first)
• NVIDIA DC GPU 全球份额 86%（2024 为 90%）

Why: When a percentage appears alone, readers easily misattribute it to the nearest category. Putting the absolute value and relative value in the same frame is the safest wording.

4.3 · Beware of "widely circulated false consensus"

Some conclusions that "everyone says" may be misunderstandings. During review, mark these categories in red:

• "H20 is banned from sale" ← It is actually a compliant version and has always been legally sold
• "the process is stuck at 7nm" ← Huawei + SMIC are already doing trial production at 5nm
• "DeepSeek training cost $5M" ← That is the operating cost; the total training cost has not been disclosed
• "NVLink bandwidth is X" ← The generation must be specified (NVLink 4 / 5 / 6)

Handling: Consensus items marked in red must be verified against primary sources such as NotebookLM. If not verified, delete them.

4.4 · Unsourced numbers must be removed or changed to qualitative statements

❌ Not allowed:

• "Huawei Ascend produces 10,000-30,000 wafers per month" (no public source)
• "NVIDIA accounts for 70% of CoWoS capacity" (precise number with no source)

✅ Must choose one:

• Remove(delete this number)
• Change to qualitative: "Huawei Ascend is constrained by lithography yield + CoWoS alternative packaging bottlenecks" (do not mention specific quantities)
• Mark as uncertain: "According to Dylan Patel's estimate, about XX-XX ten-thousand wafers" (with source + range)

Why: Research PPTs are not afraid of blank space; they are afraid of treating "it is said" as "empirical evidence." When NotebookLM answers "X does not appear in sources," that is a signal to pull more sources, not a window to pad the numbers.

Fact-checking output requirements

List a separate section in the review report ## 事实核查 · 红旗清单, mark each location that violates the 4 principles:

## 事实核查 · 红旗清单

| # | 页 | 违反原则 | 当前措辞 | 建议修法 |
|---|---|---------|---------|---------|
| 1 | P17 | 4.1 代号年份 | "Hopper 和 Blackwell 都是 900 GB/s" | Hopper NVLink 4 = 900 GB/s · Blackwell NVLink 5 = 1.8 TB/s |
| 2 | P10 | 4.2 分子分母 | "+73% YoY · DC 营收" | $681B · +73% 总营收同比 |
| 3 | P18 | 4.3 假共识 | "H20 禁售兜底" | H20 是出口管制合规版 |
| 4 | P20 | 4.4 无源数字 | "昇腾 1-3 万/月" | 去除 → 改定性"受光刻良率+CoWoS 瓶颈" |

The red-flag checklist is a mandatory review output, not an optional item. For a Research-type PPT, if any item is unaddressed, it is not allowed to enter the visual-deck stage.

Dimension 5: Redundancy detection

Which page can be removed without affecting the argument?

Go through page by page and mark:

• ✅ Core page: removing it collapses the argument
• 🔶 Supporting page: removing it weakens the argument but it can still run
• ❌ Redundant page: removing it has no impact → recommend cutting

Common redundancies in Research type:

• "Background introduction" page (if the paradox is written well, the background is automatic)
• "Methodology" page (unless the method itself is the news)
• The "summary" page of each section (if the visual anchor is done well, the summary is superfluous)

Common redundancies in Pitch type:

• "Team introduction" page placed in the middle (it should be at the end, or integrated into the pain point page)
• "Market size" page (unless the size itself is the evidence for your Ask)

Output format

After the review is complete, output the following structure:

# [PPT 标题] 故事线评审

**类型**：Research / Pitch / Teaching / Narrative
**评审时间**：YYYY-MM-DD

## 总体打分
- 结构匹配度：X/10
- 起承转合：X/10
- 关键信息承载：X/10
- 事实核查（v0.2.0 新增）：X/10 · 红旗 N 条
- 精简度：X/10

## 主要问题（按严重程度）

### 1. [问题标题 — 严重程度：高/中/低]
**现状**：[哪页 / 哪段]
**问题**：[具体描述]
**建议**：[具体改法]

### 2. ...

## 逐页评审

| 页号 | 当前内容 | 评级 | 建议 |
|-----|---------|------|-----|
| 1 | [内容] | ✅ 核心 | 保留 |
| 2 | [内容] | 🔶 支撑 | 可考虑合并到 P3 |
| 3 | [内容] | ❌ 冗余 | 建议砍 |
| ... |

## 关键锚点建议

观众会记住的页（建议上 key visual / 视觉加重）：
- 第 X 页：[为什么这里]
- 第 Y 页：[为什么这里]

## 下一步

→ 修订故事线后，跑 `visual-deck` 产出视觉 PPT
→ 如果修订后想再评审，重跑本 skill

Claude Integration

When the user triggers this skill:

1. First ask for type: if the user did not say, ask "What type of PPT is this?" (give the four options). If the type is unclear → run ppt-classify
2. Collect the storyline: have the user paste the current chapter list / page outline / draft file
3. Score against the fitting structure for the type: if the structure uses the wrong type (e.g., research but with a pitch opening),stop immediately, do not do a page-level review—tell the user "first decide whether to change the structure or change the type"
4. Go through page by page: mark according to the redundancy detection ✅/🔶/❌
5. Give key anchor suggestions: point out which pages should have a key visual
6. Produce the review report

Key attitude: review is not liking. Pointing out problems is the core value of this skill. Do not mark ❌ pages as 🔶 just because you are afraid of offending people.

Design source

This skill is distilled from EP10 "AI Makes PPT" Experiment B:

• Type-aware structural judgment ← Experiment A found that the storyline structures of pitch and research are fundamentally different; using the wrong type is a disaster
• Four fitting structures ← Comprehensive reference: pitch (Sequoia Capital pitch deck template), research (Dylan Patel / Silicon Valley 101 research format), teaching (instructional design's ability objectives + steps), narrative (Joseph Campbell's hero's journey + Robert McKee Story)
• Redundancy detection ← Borrowed from gstack plan-ceo-review 's SCOPE REDUCTION mode—cutting to the essence can expose the true core
• Fact-checking 4 principles (v0.2.0) ← EP10 dry run 2026-04-19 practical finding: during the AI argumentation stage, it fabricates technical terms ("Kyber cabinet" does not actually exist in the 43 sources). Principles distilled from 6 real corrections: code name years / numerator and denominator / beware of false consensus / unsourced numbers

See details:notes/experiment-b-three-skill-design.md · episodes/ep10-ai-ppt/notes/dryrun-2026-04-19.md