generate-prompts

name	generate-prompts
model	sonnet
description	Generate per-scene Gemini image prompts with brand styling
disable-model-invocation	false
argument-hint	<book-slug>

Generate Image Prompts

Create per-scene AI image prompts in Gemini format for a Bookie book video. Combines the storyboard's visual concepts with actual voice timing and brand styling into self-contained prompts ready to paste into Gemini.

Context

• When to use: After make voice produces actual timing. Before make scenes.
• Input: storyboard.md + chunks-display.md + section-timing.json + style-guide.md
• Output: image-prompts.md (per-scene Gemini prompts with timing)

Pipeline Position

make voice → section-timing.json (actual timing)
        ↓
/generate-prompts → image-prompts.md  ← YOU ARE HERE
        ↓
make scenes → generated images

Steps

1. Validate: Check that $ARGUMENTS is provided. If missing, ask Hai for the book slug. Set SLUG=$ARGUMENTS. Check that projects/ai-book-video/books/$SLUG/ exists.

2. Read input files:

   • Read projects/ai-book-video/books/$SLUG/storyboard.md — visual concepts per scene
   • Read projects/ai-book-video/books/$SLUG/chunks-display.md — narration text per scene
   • Read projects/ai-book-video/books/$SLUG/output/section-timing.json — actual scene timing from voice
   • Read projects/ai-book-video/brand/style-guide.md — brand colors and style rules

   If section-timing.json is missing, warn: "Voice not generated yet. Run make voice BOOK=$SLUG first, or proceed without timing?" If proceeding without timing, omit timestamp fields.

3. Build style prefix: From style-guide.md, construct the Gemini style instruction that prefixes every prompt. Include:

   • Art direction (textured editorial illustration — risograph-inspired grain, subtle paper texture, visible brushwork, soft shadows, layered depth. Minimal, metaphor-driven.)
   • Brand colors (green #368C06 for growth/positive, orange #C86108 for tension/emphasis)
   • Background color (#FAFDF5)
   • Constraints: "no text, no watermark, no realistic faces"

   The style guide is the single source of truth for colors — read it fresh each time.

4. Detect mode: If image-prompts.md exists → ask whether to overwrite or revise specific scenes. If missing → new draft.

5. Generate prompts: For each scene:

   • Pull visual concept from storyboard.md
   • Pull narration excerpt from chunks-display.md (key line for context)
   • Pull timestamp from section-timing.json (startMs → endMs, converted to M:SS)
   • Write a Gemini prompt: style prefix + visual description in natural language
   • Suggest layer type: Flat (static image) or Ken Burns (slow pan/zoom)

6. Present to Hai: Show prompts for review. He may want to:

   • Change metaphors for specific scenes
   • Adjust composition
   • Merge or split scenes visually
   • Change layer types

7. Write output: Write to projects/ai-book-video/books/$SLUG/image-prompts.md

Output Format

# Image Prompts: [Book Title] — [Angle]

> Generated by /generate-prompts on [date]
> Scenes: [N] | Source timing: section-timing.json

## Style Prefix

> Prepend this to every Gemini prompt:

[Style instruction block — textured editorial, brand colors, constraints]

---

## Prompts

### Scene 01 — HOOK (0:00 - 0:12)
- **Duration**: 12.5s
- **Script**: "[Key narration line from this scene]"
- **Visual concept**: [From storyboard]
- **Prompt**: "[Complete self-contained Gemini prompt including style prefix]"
- **Layers**: Ken Burns (slow zoom out)
- **Status**: [ ] Generated  [ ] Approved

### Scene 02 — CONTEXT (0:12 - 0:22)
- **Duration**: 9.8s
- **Script**: "[Key line]"
- **Visual concept**: [From storyboard]
- **Prompt**: "[Complete prompt]"
- **Layers**: Flat
- **Status**: [ ] Generated  [ ] Approved

...

## Summary

| Scene | Label | Duration | Layers | Status |
|-------|-------|----------|--------|--------|
| 01 | HOOK | 12.5s | Ken Burns | [ ] |
| 02 | CONTEXT | 9.8s | Flat | [ ] |
...

Prompt Writing Guidelines

Every prompt is self-contained — paste it into Gemini without any other context. This means each prompt includes the style instructions, not just a reference to them.

Focus on metaphor, not literal depiction. Instead of "a person reading a book," try "an open book floating in space, its pages transforming into glowing pathways." Textured editorial illustration — NOT flat vector, NOT stock photography. Every image should feel like a premium magazine print: grain, depth, visible texture.

Visual principles:

• No text in images — the subtitle handles text
• No realistic faces — use simple character silhouettes or abstract figures
• Brand green (#368C06) for positive/growth elements
• Brand orange (#C86108) for tension/emphasis/warning
• Light background (#FAFDF5) as base
• Consistent style across all scenes — same textured editorial quality, same grain level

Gemini-specific:

• Natural language descriptions (not tags or keywords)
• Include "no text, no watermark" inline in the prompt
• Describe composition: foreground/background, lighting, perspective
• Mention color palette explicitly for key elements

Layer suggestions (feeds into generate-scenes.sh for per-scene visual override):

• Ken Burns (slow zoom in): Zoom into details — intimacy, focus, revelation
• Ken Burns (slow zoom out): Zoom out to reveal — context, scale, perspective shift
• Ken Burns (slow pan left to right): Horizontal motion — progression, journey, transition
• Ken Burns (slow pan right to left): Reverse motion — looking back, contrast
• Flat: Static image — diagrams, comparisons, focused concepts, data

Important

• Timestamps come from section-timing.json (actual voice timing), not estimates
• The style guide is the single source of truth for brand colors — read it, don't hardcode
• All non-prompt text in Vietnamese
• Image prompts themselves can be in English (Gemini handles EN better for image generation)