AI cannot write high-quality content for the user. It can only help the user express their own content better.

Step 1 is ALWAYS: collect the user's original words. Their transcripts, their articles, their notes, their voice. AI-generated content without user source material is garbage — polished, plausible, unusable garbage.

Weight hierarchy: User's own words > User-approved external material > AI synthesis > AI invention.

The output must sound like the user, at their best. Something they could actually say, confidently, in their own voice.

references/content-creation-first-law.md

• Transcripts of their past talks, meetings, or discussions
• Articles they have written or approved
• Notes or drafts they have prepared
• Previous decks they have delivered
• Voice memos or any recorded thoughts

• Search for relevant external materials (articles, reports, references)
• Present findings to the user for selection — do not assume relevance
• Only include materials the user explicitly approves

• Topic: What's the talk about? (1 sentence)
• Audience: Who's listening? (technical level, role, context)
• Duration: How long is the talk?
• Key messages: What must they remember? (3 max)
• Tone: Educational? Persuasive? Provocative? Inspirational?
• Existing content: Articles, transcripts, notes, previous decks?
• Constraints: Must-use content? Avoid topics? Brand guidelines?

1. Opening strategy: Shock? Story? Question? Demo?
2. The arc: Where's the tension? Where's the release?
3. Transition logic: How does each slide lead to the next?
4. The "one thing": If they forget everything, what's the ONE thing?
5. Call to action: What do they do after the talk?

• ❌ Too many slides for the time (crowded, rushed)
• ❌ Jumping into details without setting context
• ❌ No emotional arc (flat, forgettable)
• ❌ Ending without a clear takeaway
• ❌ Trying to teach too many things at once

1. ...
2. ...
3. ...

• Original user prompt saved
• Existing articles/notes/transcripts
• Previous decks

1. [Point 1]
2. [Point 2]
3. [Point 3]

00-上游/

<STYLE_INSTRUCTIONS>
Design Aesthetic: [Description]

Background:
  Texture: [clean/grid/organic/etc]
  Base Color: [#HEX]

Typography:
  Headlines: [Style, size, color, weight]
  Body: [Style, size, color, weight]

Color Palette:
  Primary Text: [#HEX] - usage
  Body Text: [#HEX] - usage
  Background: [#HEX]
  Accent 1: [#HEX] - usage
  Accent 2: [#HEX] - usage
  Accent 3: [#HEX] - usage

Visual Elements:
  - [Element 1]
  - [Element 2]

Density Guidelines:
  - Max [N] text elements per slide
  - [Other rules]

Style Rules:
  Do: [List]
  Don't: [List]
</STYLE_INSTRUCTIONS>

content.md

1. User's own words (from Phase 0 source materials)
2. User-approved external references
3. User's explicit statements during Phase 1 discussion

[TODO: user to provide]

/baoyu-slide-deck 00-上游/content.md --prompts-only [--style <preset>]

/

Skill({"skill": "baoyu-slide-deck", "args": "00-上游/content.md --prompts-only [--style <preset>]"})

1. Inject narrative-brief: Append

   narrative-brief.md

   (or its ABCDEFG arc section) to the top of

   content.md

   so baoyu receives the narrative structure, not just raw content.
2. Inject confirmed choices: Prepend a metadata block to

   content.md

   with the already-confirmed choices. This acts as a strong signal for baoyu's auto-detect and reduces the chance of style drift during confirmation:
   markdown

   <!-- CONFIRMED CHOICES — do not override without discussion -->
   - Style: [preset name or custom]
   - Audience: [audience from Phase 1]
   - Slide count: [N slides for X-min talk]
   - Language: [zh/en/etc]
   - Review preference: [skip outline / skip prompts / none]

3. Style selection: If user specified a baoyu preset → use it; if custom

   style-instructions.md

   exists → use

   --style custom

   ; otherwise auto-detect.

vector-illustration

bold-editorial

hand-drawn-edu

chalkboard

sketch-notes

corporate

minimal

bold-editorial

editorial-infographic

dark-atmospheric

pixel-art

watercolor

blueprint

intuition-machine

notion

fantasy-animation

scientific

vintage

prompts/

1. Copy to user's structure: Move/copy prompts to

   03-prompts/

2. Inject custom style: If

   style-instructions.md

   exists, ensure FULL content is embedded in every prompt file
3. Add narrative goal: Append

   // NARRATIVE GOAL

   section to each prompt based on

   narrative-brief.md

---

1. Display prompt summary table
2. Ask user: "Ready to generate images?"
3. If edits needed → user edits

   03-prompts/*.md

   → regenerate specific prompts via baoyu-slide-deck

1. Test slide: Read

   01-slide-cover.png

   (or first generated slide)
2. Style check: Compare against

   style-instructions.md

3. Text check: Verify Chinese/English text legibility
4. If issues: Update affected

   03-prompts/*.md

   → copy

   cp -r 03-prompts prompts

   → regenerate via

   /baoyu-slide-deck . --regenerate N

   (或

   Skill({"skill": "baoyu-slide-deck", "args": ". --regenerate N"})

   ) → clean up

   rm -rf prompts

slide-deck/{topic-slug}/

slide-deck/{topic-slug}/
├── source-{slug}.md
├── outline.md
├── prompts/
├── *.png
├── {topic-slug}.pptx
└── {topic-slug}.pdf

{project-name}/
├── 00-上游/                    # Source materials
│   ├── prompt-最初提示词.txt   # Original user prompt (if saved)
│   ├── narrative-brief.md      # Phase 1 output
│   ├── content.md              # Phase 2 output (baoyu input)
│   ├── style-instructions.md   # Visual design SSOT
│   └── outline.md              # From baoyu-slide-deck
├── 01-成品/                    # Final deliverables
│   ├── {project-name}.pdf
│   └── {project-name}.pptx
├── 02-slides/                  # Generated PNGs (当前版本)
│   ├── 01-slide-cover.png
│   └── ...
├── 03-prompts/                 # Per-slide prompts (SSOT)
│   ├── v6/                     # 支持版本子目录（如 v6, v7...）
│   │   ├── 01-slide-cover.md
│   │   └── ...
│   └── 01-slide-cover.md       # 或平铺结构
├── speaker-notes.md            # Auto-extracted from 03-prompts/ via extract_notes.py
├── v6/                         # baoyu-slide-deck 临时输出（需搬迁到 02-slides/）
│   └── ...
└── _archive/                   # Historical versions
    └── v1/

• 03-prompts/

  支持平铺或版本子目录（

  v6/

  ,

  v7/

  ）。当同一项目多次迭代时，用子目录保留历史版本。
• baoyu-slide-deck 可能直接输出到项目根目录的临时文件夹（如

  v6/

  ）。Post-processing 时需将这些 PNG 移动到

  02-slides/

  。

uv run scripts/archive_version.py --project /path/to/project

02-slides/

03-prompts/

_archive/v{N}/

scripts/extract_notes.py

03-prompts/*.md

uv run scripts/extract_notes.py --prompts 03-prompts --output speaker-notes.md

• // NARRATIVE GOAL

  sections

• // SPEAKER NOTES

  sections
• Falls back to

  // KEY CONTENT

  if neither found

speaker-notes.md

undefined

• Save as

  00-上游/prompt-最初提示词.txt

• PDF opens and all slides render correctly
• PPTX opens without errors
• PNG sequence numbered correctly (01, 02, ...)
• Speaker notes cover all slides
• Style consistent across all slides
• No garbled or missing text

• 00-上游/

  contains all source SSOT files

• 03-prompts/

  contains all prompt files

• references/content-creation-first-law.md

  — Universal principle: user's voice is primary, applies to all content types (slides, articles, ads, courses)

• references/narrative-design-guide.md

  — ABCDEFG model detailed guide

• references/prompt-templates/

  — Prompt templates for common slide types

• references/style-gallery.md

  — Visual style gallery with examples