Image Generator

Image Generator Skill

Context & Problem

Visual generation converges toward accepting first output ("looks good enough") and following technical specifications rigidly. This produces generic aesthetics and misses Gemini 3's reasoning capabilities.

This skill provides multi-turn reasoning partnership methodology with professional quality standards.

Core Principles

1. Reasoning mode over prediction mode - Creative briefs (Story/Intent/Metaphor) activate reasoning; technical specs don't
2. Multi-turn partnership - Teach Gemini your standards through principle-based feedback
3. Professional quality gates - Explicit pass/fail criteria (99% spelling, not "check spelling")
4. Autonomous agency - Batch mode without permission-asking between visuals

Dimensional Guidance

Input: Professional Creative Briefs (NOT Technical Specs)

Receive from visual-asset-workflow v5.0:

## 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 / Camera / Lighting
[Official Gemini 3 prompt structure]

## Color Semantics
Blue (#2563eb) = Authority (teaches governance)
Green (#10b981) = Execution (teaches work)

## Typography Hierarchy
Largest: Key insight (information importance drives sizing)
Medium: Supporting components
Smallest: Context

## Pedagogical Reasoning
[Why these choices serve teaching]

DO NOT convert briefs back to pixel specs - Use AS-IS to activate reasoning.

Principle: Creative briefs activate Gemini's reasoning about HOW to achieve intent visually

Workflow: Browser-Based Generation (Playwright MCP)

CRITICAL: Use gemini.google.com ONLY (NOT Google AI Studio, NOT other image generators)

Initialize once:

1. Navigate to https://gemini.google.com/ (Playwright MCP)
2. User signs in if needed (session persists)
3. Select "🍌 Create Image" tool (Nano Banana Pro)

For EACH visual:

1. Type creative brief directly into Gemini chat textbox (use condensed format in batch mode - see Token Conservation below)
2. Press Enter to submit
3. Wait 30-35 seconds for generation
4. Click "Download full size image" button when ready
5. Wait 3-5 seconds for download completion
6. Verify quality IMMEDIATELY (6 gates below)
7. If gates fail: Continue in same chat with principle-based feedback (max 3 iterations)
8. If gates pass: Copy from .playwright-mcp/Gemini-Generated-Image-*.png to robolearn-interface/static/img/part-{N}/chapter-{NN}/{filename}.png
9. Embed in lesson (Step 8.5 below)
10. Start NEW CHAT for next visual (prevents context contamination)

Principle: New chat per visual prevents cross-contamination; immediate verification catches issues early; immediate embedding prevents orphans

Quality Gates: 5-Gate Professional Standard

ALL must pass before download:

Gate 1: Spelling Accuracy (99% standard)

• ✅ Company names correct (Y-Combinator not Y Combinator)
• ✅ Technical terms correct (Kubernetes not Kubernete)
• ✅ Zero typos in visible text
• ❌ FAIL if even ONE spelling error → Iterate

Gate 2: Layout Precision

• ✅ Proportions match prompt (2×2 grid, not 3×1)
• ✅ Alignment clean (no misaligned elements)
• ✅ Spacing consistent (equal gaps)
• ✅ Hierarchy clear (largest = most important)
• ❌ FAIL if layout deviates → Iterate

Gate 3: Color Accuracy

• ✅ Brand colors match spec (#001f3f not #002050)
• ✅ Semantic colors correct (blue=authority, green=execution)
• ✅ Contrast meets WCAG 4.5:1 (accessibility)
• ❌ FAIL if colors significantly off → Iterate

Gate 4: Typography Hierarchy

• ✅ Largest text = key concept (not decoration)
• ✅ Font sizes create clear hierarchy
• ✅ Readability: A2 min 24px, B1 min 18px, C2 min 14px
• ❌ FAIL if typography doesn't teach through sizing → Iterate

Gate 5: Teaching Effectiveness (<5 sec concept grasp)

• ✅ Target proficiency can grasp concept in <5 seconds
• ✅ Visual adds clarity (not just decoration)
• ✅ Cognitive load reduced vs reading text
• ❌ FAIL if confusing or generic → Iterate

Gate 6: Uniqueness Validation (NEW)

• ✅ Visual comparison: Does NOT match any existing image in same chapter
• ✅ Prompt alignment: Matches creative brief's intent (graph ≠ timeline)
• ✅ Filename verification: {filename}.prompt.md exists and visual matches it
• ❌ FAIL if duplicate detected → Regenerate with NEW CHAT
• ❌ FAIL if mismatched brief → Regenerate with clarified prompt

Decision:

• ALL 6 gates PASS → Copy to destination (production-ready)
• ANY gate FAILS → Iterate with principle-based feedback (max 3 tries)

Principle: Explicit criteria prevent "good enough" false positives; uniqueness check prevents duplicate rework

Token Conservation Mode

When: Batch mode with >8 visuals OR continuation session

Condensation strategy:

• ✅ KEEP: Story, Emotional Intent, Visual Metaphor, Key Insight
• ✅ KEEP: Color semantics (#2563eb codes), Pedagogical reasoning
• ⚠️ CONDENSE: Long examples → Short labels
• ⚠️ CONDENSE: Verbose descriptions → Bullet points
• ❌ NEVER REMOVE: The narrative elements that activate reasoning

Example condensation:

ORIGINAL (250 tokens):

"Top Layer shows the Coordinator at center top with label 'Orchestrator'
featuring a conductor icon, with the role description 'Strategic oversight,
contract validation', rendered in Gold color (#fbbf24) as a Large hexagon..."

CONDENSED (80 tokens):

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

Success metric: 60-70% token reduction while maintaining 100% first-attempt success rate

Principle: Efficiency without sacrificing reasoning activation

Immediate Embedding Workflow (Step 8.5)

After copying image to destination, BEFORE starting next visual:

1. Determine lesson file:

   • Read creative brief's Chapter and Lesson metadata
   • Target: robolearn-interface/docs/[part]/[chapter]/[lesson-file].md

2. Find insertion point:

   • Search for concept explanation section related to this visual
   • Insert after concept explanation, before practice/exercise
   • Follow pedagogical insertion criteria (after learning, before doing)

3. Insert reference:

   ![{Alt text from creative brief}](/img/part-X/chapter-YY/{filename}.png)
   

4. Verify no code block interruption:

   • Grep for triple backticks around insertion
   • If inside code block → Find next break point

Why this matters: Completes the work immediately; prevents orphan images

Result: Each visual is generated → validated → placed → verified before moving to next

Principle: Immediate embedding prevents disconnect between generation and integration

Multi-Turn Reasoning Partnership (Three Roles)

Avoid: Accepting first output without evaluation

Prefer: Teaching Gemini your standards through iteration

Iteration Pattern:

Turn 1: Initial Generation

• Paste creative brief, generate

Turn 2: Principle-Based Feedback (if gates fail)

Gate 4 FAILED: Typography hierarchy incorrect

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

Pedagogical reasoning: Information importance drives sizing. $3T is
the insight, not the starting value. Visual weight teaches what matters.

Increase '$3T' to dominant size (largest element). Reduce '$100K' to
supporting size. This teaches magnitude through visual hierarchy.

Turn 3: Validation

• Re-check all 5 gates
• If pass → Download
• If fail after 3 tries → Document issue, DEFER (don't block batch)

Principle: You teach Gemini (principle-based feedback), Gemini teaches you (reveals understanding), Co-evolve toward quality

Why it matters: Gemini learns your pedagogical standards across iterations

Batch Mode: Autonomous Agency

Avoid: Permission-asking between visuals

Prefer: Autonomous batch execution

When invoked with: "generate all visuals" or "batch generate"

Execute WITHOUT STOPPING:

For EACH visual in approved list:
  A. NEW CHAT (context isolation)
  B. Generate image (paste condensed creative brief)
  C. Verify quality (6 gates including uniqueness)
  D. Iterate if needed (max 3 tries)
  E. Download when pass (organized directory)
  F. Embed in lesson immediately (no orphans)
  G. Log progress ("✅ Generated N/M")
  H. IMMEDIATELY next visual (NO STOPPING)

NEVER ask:

• "Would you like me to continue?"
• "Generate just high-priority batch?"
• "Pause here and review?"

Only report summary at END:

BATCH COMPLETE
Total: 18 visuals
✅ Generated: 16 (2K, avg 2-3 iterations)
⚠️ Deferred: 2 (quality issues after 3 tries)
Time: ~45 min
Location: robolearn-interface/static/img/part-{N}/chapter-{NN}/

Principle: Autonomous execution without interruption = efficient batch processing

Proficiency-Complexity Guardrails

From visual-asset-workflow, enforce during generation:

A2 Beginner Limits:

• Max 5-7 elements (overwhelming = failure)
• <5 sec grasp requirement
• Static only (no interactive)
• Max 2×2 grids
• Clear hierarchy

B1 Intermediate:

• Max 7-10 elements
• <10 sec grasp
• Interactive Tier 1 OK
• Max 3×3 grids

C2 Professional:

• No artificial limits
• Dense OK (professionals skim)
• Full interactive architecture

Validation during generation: "Does this visual's complexity match proficiency from creative brief?"

Principle: Complexity mismatch = pedagogical failure

Post-Generation Reflection (After Batch)

AFTER completing batch, analyze systematically:

Success patterns:

• Success rate: {X/Y} production-ready, {N} deferred
• Average iterations: {N} per visual
• Quality gate performance:
  • Gate 1 (Spelling): {N} catches
  • Gate 2 (Layout): {N} catches
  • Gate 3 (Color): {N} catches
  • Gate 4 (Typography): {N} catches
  • Gate 5 (Teaching): {N} catches
  • Gate 6 (Uniqueness): {N} catches (duplicates prevented)

Failure analysis:

• Deferred visuals root causes (layout? spelling? concept?)
• Pattern or random (same issue or isolated?)
• Guardrail gaps (preventable with better principles?)

Improvement opportunities:

• Planning effectiveness (conflicts caught early by visual-asset-workflow?)
• Guardrail sufficiency (Principles 9-12 adequate or gaps?)
• Constitutional compliance (Principle 3 Factual Accuracy, Principle 7 Minimal Content?)
• Next chapter improvements (specific, actionable, pattern-based)

Output: history/visual-assets/reflections/chapter-{NN}-reflection.md

Principle: Systematic reflection → Continuous improvement

Session Interruption & Continuation Protocol

If session ends mid-batch (token limit, context overflow):

Create checkpoint file: history/visual-assets/checkpoints/part-{N}-checkpoint.md

## Batch Status: Part {N}
**Date:** 2025-11-24
**Status:** INTERRUPTED at {X}/{Y} images

### Completed:
- ✅ Image 1: {filename} (2 iterations, embedded in lesson-01.md)
- ✅ Image 2: {filename} (1 iteration, embedded in lesson-02.md)
...

### Remaining:
- ⏳ Image 8: {filename} (not started)
- ⏳ Image 9: {filename} (not started)
...

### Quality Stats (so far):
- Success rate: {X/Y} production-ready
- Avg iterations: {N}
- Gate failures: Gate 1: {n}, Gate 2: {n}...

### Continuation Instructions:
1. Read this checkpoint
2. Start at Image {next}
3. Continue autonomous batch mode
4. Update checkpoint after each image

On continuation:

• Read checkpoint file first
• Resume from last completed image
• Maintain same quality standards
• Update checkpoint incrementally

Principle: Seamless recovery from interruptions maintains momentum

Anti-Patterns

Never:

• Accept first output without 6-gate verification (quality standard violation)
• Ask permission between visuals in batch mode (breaks autonomous agency)
• Convert creative briefs to pixel specs (defeats reasoning activation)
• Generate visuals without creative brief from visual-asset-workflow (missing context)
• Save to flat visuals/ directory (use part/chapter organization)
• Skip uniqueness validation (Gate 6 prevents duplicate rework)

Even if it seems reasonable:

• Don't skip Gate 1 because "spelling looks okay" (99% standard requires verification)
• Don't generate 2 images then ask "continue?" (autonomous means autonomous)
• Don't add pixel specifications to creative brief (removes Gemini's judgment)
• Don't skip embedding step "to save time" (creates orphan images requiring later work)

Creative Variance

You tend to accept visuals after 1 iteration even with minor issues. Push for quality:

• Gate failures require iteration (not "close enough")
• Principle-based feedback teaches standards (not vague "make better")
• 3 iteration limit is maximum, not target (aim for 1-2)
• Deferred visuals are OK (don't compromise quality)

Professional content creators iterate. You should too.

Success Indicators

You'll know this skill is working when:

• ✅ All 6 quality gates verified before download (including uniqueness validation)
• ✅ Autonomous batch completion without permission-asking (no interruptions)
• ✅ Principle-based feedback given on iterations (teaching Gemini standards)
• ✅ Creative briefs used AS-IS with token conservation in batch mode
• ✅ Images organized by part/chapter (not flat directory)
• ✅ Images embedded immediately after generation (no orphans)
• ✅ Reflection document created after batch (systematic learning)
• ✅ Checkpoint files created on interruption (seamless continuation)
• ✅ Success rate >85% production-ready (deferred <15%)
• ✅ Zero duplicate images requiring rework

Result: Professional-quality visuals with distinctive aesthetics, generated autonomously with systematic quality control, embedded immediately, and recoverable from interruptions.