references/iterative_refinement.md
Smart Iterative Refinement
The generate-review-refine loop, the command-line reference, and configuration.
Smart Iterative Refinement
The diagram below shows the successful review path. If generation fails, the loop stops without replaying the paid call; the latest saved draft is retained. If review produces no usable score or verdict, the loop stops with quality unverified. A verdict requesting changes can trigger refinement even without a score. A saved PNG is not a factual check.
How It Works
Generate PNG with Nano Banana 2
-> Review with Gemini 3.7 Flash
-> Score meets target AND no improvements requested?
YES: retain PNG, quality_met=true
NO: refine prompt while budget remains
(otherwise retain draft, quality_met=false)
Quality Review Criteria
Gemini 3.7 Flash evaluates each infographic on:
-
Visual Hierarchy & Layout (0-2 points)
- Clear visual hierarchy
- Logical reading flow
- Balanced composition
-
Typography & Readability (0-2 points)
- Readable text
- Bold headlines
- No overlapping
-
Data Visualization (0-2 points)
- Prominent numbers
- Clear charts/icons
- Proper labels
-
Color & Accessibility (0-2 points)
- Professional colors
- Sufficient contrast
- Colorblind-friendly
-
Overall Impact (0-2 points)
- Professional appearance
- Free of visual bugs
- Achieves communication goal
Review Log
Each generation produces a JSON review log:
{
"user_prompt": "5 benefits of exercise...",
"infographic_type": "list",
"style": "healthcare",
"doc_type": "marketing",
"quality_threshold": 8.5,
"iterations": [
{
"iteration": 1,
"image_path": "figures/exercise_v1.png",
"score": 8.7,
"needs_improvement": false,
"critique": "SCORE: 8.7\nSTRENGTHS:..."
}
],
"final_score": 8.7,
"quality_met": true,
"termination_reason": "quality_met",
"image_model": "google/gemini-3.1-flash-image",
"review_model": "google/gemini-3.7-flash",
"early_stop": true,
"early_stop_reason": "Quality score 8.7 meets threshold 8.5"
}
Command-Line Reference
python skills/infographics/scripts/generate_infographic.py [OPTIONS] PROMPT
Arguments:
PROMPT Description of the infographic content
Options:
-o, --output PATH Output .png path (required)
-t, --type TYPE Infographic type preset
-s, --style STYLE Industry style preset
-p, --palette PALETTE Colorblind-safe palette
-b, --background COLOR Background color (default: white)
--doc-type TYPE Document type for quality threshold
--iterations N Positive generation-attempt budget (default: 3)
-r, --research Gather candidate facts with Sonar Pro
--context-image PATH Reference image; repeat up to 14 times
--api-key KEY OpenRouter API key
-v, --verbose Verbose output
--list-options List all available options
success means a draft was saved. quality_met is true only for a usable score at or
above the target with no requested improvements. termination_reason is quality_met,
review_unavailable, max_iterations, or generation_failed. Even a passing score
requires source, spelling, number, geometry, contrast, and accessibility checks.
--context-image supports local PNG, JPEG, GIF, and WebP references, sent as base64
input_references to the Image API. Up to 14 references are supported by the currently
reviewed Gemini providers; the script sends one output request per iteration. Generated
PNG dimensions are provider-selected: a requested pixel size written in a prompt is not
an API-enforced size. Inspect the actual artifact before print or platform delivery.
List All Options
python skills/infographics/scripts/generate_infographic.py --list-options
Configuration
API Key Setup
Set your OpenRouter API key:
export OPENROUTER_API_KEY='your_api_key_here'
Get an API key at: https://openrouter.ai/keys