image-generator

Image Generator

Generate professional teaching visuals using Gemini 3 with multi-turn reasoning partnership.

Quick Start

# 1. Start browser (via browser-use skill)
bash .claude/skills/browser-use/scripts/start-server.sh

# 2. Navigate to Gemini
# Use browser_navigate to https://gemini.google.com/

# 3. Generate image from creative brief
# Paste creative brief → Wait 30-35s → Verify 6 gates → Download

Core Principles

1. Reasoning over prediction - Creative briefs (Story/Intent/Metaphor) activate reasoning; pixel specs don't
2. Multi-turn partnership - Teach Gemini your standards through principle-based feedback
3. 6-gate quality - Explicit pass/fail before download
4. Autonomous batch - No permission-asking between visuals

Input: Creative Brief Format

Receive from visual-asset-workflow:

## The Story
[Narrative about what's visualized]

## Emotional Intent
[What it should FEEL like]

## Visual Metaphor
[Universal concept for instant comprehension]

## Subject / Composition / Action / Location / Style
[Gemini 3 prompt structure]

## Color Semantics
Blue (#2563eb) = Authority | Green (#10b981) = Execution

## Typography Hierarchy
Largest: Key insight | Medium: Supporting | Smallest: Context

Do NOT convert to pixel specs - use as-is to activate reasoning.

Workflow (Per Visual)

Quality Gates (ALL Must Pass)

Decision: ALL pass → Download | ANY fail → Iterate (max 3 tries)

Iteration: Principle-Based Feedback

When gate fails, provide teaching feedback:

Gate 4 FAILED: Typography hierarchy incorrect

The largest text is "$100K" (supporting detail) but should be "$3T"
(key insight students must grasp).

Increase '$3T' to dominant size. Reduce '$100K' to supporting size.
Information importance drives sizing.

Batch Mode

When invoked with "generate all visuals":

For EACH visual in list:
  A. NEW CHAT (context isolation)
  B. Generate (paste brief)
  C. Verify 6 gates
  D. Iterate if needed (max 3)
  E. Download when pass
  F. Embed in lesson
  G. Log "✅ N/M"
  H. NEXT (no stopping)

Never ask: "Continue?" "Pause here?" "Review?"

Report at END only:

BATCH COMPLETE
✅ Generated: 16/18
⚠️ Deferred: 2 (quality issues)
Location: apps/learn-app/static/img/part-{N}/

Proficiency Limits

Token Conservation (Batch Mode)

For >8 visuals, condense briefs:

Original (250 tokens):

"Top Layer shows Coordinator at center top with label 'Orchestrator'
featuring conductor icon, with role 'Strategic oversight'..."

Condensed (80 tokens):

"Top Layer - Coordinator: Center top, 'Orchestrator' (conductor),
Role: 'Strategic oversight', Gold (#fbbf24), Large hexagon."

Keep: Story, Intent, Metaphor, Colors, Reasoning Condense: Long examples → Short labels

Anti-Patterns

Session Interruption

If session ends mid-batch, create checkpoint:

# Checkpoint: Part {N}
Status: INTERRUPTED at 8/18

## Completed:
- ✅ Image 1: filename (embedded lesson-01.md)
- ✅ Image 2: filename (embedded lesson-02.md)

## Remaining:
- ⏳ Image 8: filename

On continuation: Read checkpoint → Resume → Update incrementally

Success Indicators

• ✅ All 6 gates verified before download
• ✅ Batch completion without permission-asking
• ✅ Principle-based iteration feedback
• ✅ Images organized by part/chapter
• ✅ Immediate embedding (no orphans)
• ✅ >85% production-ready rate