Nano Banana

Unified image generation workflow using a fixed Python script with JSON configuration. Eliminates AI hallucinations by avoiding dynamic code generation.

Two modes:

Unified Generation: detailed prompt → write JSON config → run generate_images.py
Interactive Prompting: vague prompt or "help me craft" → guide user, then generate

When to Use

Image generation ("draw", "create", "generate"), slides, presentations
Image editing with AI
Prompt help ("help me craft", "improve my prompt")
Brand styles ("use trendlife style", "notebooklm style")
Reproducible generation ("use seed 42", "regenerate that image")

Mode selection

Interactive Prompting: prompt help requested OR prompt is <5 words
Unified Generation: detailed prompt provided

🚨 CRITICAL: API Routing

The script picks the API based on IMAGE_GEN_MODEL and reference_image:

Condition	API	Endpoint
`gpt-image-2`, no `reference_image`, no TrendLife featured	`client.images.generate()`	POST `/images/generations`
`gpt-image-2` + `reference_image`	`client.images.edit()`	POST `/images/edits`
`gpt-image-2` + `style: "trendlife"` + `layout: "featured"`	`client.images.edit()` (logo as image)	POST `/images/edits`
Gemini or other	`client.chat.completions.create()`	POST `/chat/completions`

gpt-image-2: Always uses quality="low" — other values timeout or return 400. Seed and temperature unsupported (silently ignored, one warning per job).

Gemini (gemini-3-pro-image, gemini-3.1-flash-image): use chat.completions with extra_body={"response_modalities": ["IMAGE"]}.

client.chat.completions.create(
    model=os.environ.get("IMAGE_GEN_MODEL") or "gemini-3-pro-image",
    messages=[{"role": "user", "content": prompt}],
    seed=seed,
    temperature=temperature,
    extra_body={"response_modalities": ["IMAGE"]},
)

IMAGE_GEN_MODEL: NEVER override

If IMAGE_GEN_MODEL is set, use it EXACTLY as-is. Custom endpoints have their own model names — do not append suffixes like -preview. Only choose a default when the env var is unset.

model = os.environ.get("IMAGE_GEN_MODEL")
if not model:
    model = "gemini-3-pro-image"
# Use as-is — do NOT add suffixes

Full model × feature matrix: references/api-compatibility.md

Unified Generation Workflow

Create config — write JSON to system temp dir (NOT skill directory):
- Write tool → {temp_dir}/nano-banana-config-{timestamp}.json
Execute — uv run --managed-python ${CLAUDE_SKILL_DIR}/scripts/generate_images.py --config <temp_config_path>
- CRITICAL: use ${CLAUDE_SKILL_DIR} prefix without cd — keeps cwd in user's project so relative paths in config resolve correctly.
- ❌ cd scripts && uv run --managed-python generate_images.py ...
- ✅ uv run --managed-python ${CLAUDE_SKILL_DIR}/scripts/generate_images.py ...
Track progress — monitor progress/results files (background tasks)
Return paths — report generated image locations

5+ slides → automatic background execution. Details: references/batch-generation.md

Config

Minimal:

{
  "slides": [{"number": 1, "prompt": "...", "style": "trendlife"}],
  "output_dir": "./001-feature-name/"
}

output_dir MUST be ./NNN-short-name/ (relative path, 3-digit number, kebab-case)
No absolute paths, no model field (use env var), no seed/temperature for gpt-image-2

Full schema, format selection, layout field rules, gpt-image-2 examples: references/config-reference.md

Seed & Temperature (Gemini only)

"seed 42" → "seed": 42
"regenerate with seed 392664860" → use seed from previous results JSON
"temperature 0.5" / "temp 1.5" / "more creative" → "temperature" field
Priority: slide.seed > config.seed > auto-generate

Same seed + same prompt + same temperature = identical image. Auto-seeds are recorded in {output_dir}/generation-results.json.

Full parsing patterns + reproducibility detail: references/temperature-seed.md

Style Detection

Detect style: "trendlife", style: "notebooklm", or natural language ("use trendlife style").

TrendLife: Trend Red (#D71920), automatic logo overlay — see references/logo-overlay.md and references/brand-styles.md
NotebookLM: clean presentation aesthetic. ⚠️ NEVER use "NotebookLM" name/logo in prompts (trademark) — see references/slide-deck-styles.md

Priority: inline spec → ask in Interactive Mode → no style (default).

TrendLife layout field

Value	Use for	Logo handling
`"featured"`	title, divider, closing	Logo as reference image, AI integrates naturally
`"content"`	information slides	Post-generation overlay, bottom-right corner
omitted	(auto-detect from keywords — less reliable)	—

Always specify layout explicitly for TrendLife. Full detection rules + manual override: references/logo-overlay.md

Interactive Prompting Mode

Trigger: prompt help requested OR prompt <5 words.

Step	Action
1. Gather	Check for reference images, existing prompts, inline style specs
2. Clarify	`AskUserQuestion`: output type? subject? style?
3. Select technique	Pick from 16+ patterns in `references/guide.md`
4. Generate prompt	Apply technique + brand style + aspect ratio
5. Present	Show prompt with explanation and variations
6. Execute	Run via Unified Generation Workflow

Brand style integration:

TrendLife (style: "trendlife"): see references/brand-styles.md
NotebookLM (style: "notebooklm"): use "clean professional presentation aesthetic" — never the "NotebookLM" name. See references/slide-deck-styles.md

Techniques and examples: references/guide.md

References

Topic	File
API compatibility, model × feature matrix	`references/api-compatibility.md`
Batch / multi-slide generation	`references/batch-generation.md`
TrendLife brand specs	`references/brand-styles.md`
Config schema, layout field, format selection	`references/config-reference.md`
Example workflows (4 scenarios)	`references/examples.md`
Prompting techniques (16+)	`references/guide.md`
Logo overlay rules + manual override	`references/logo-overlay.md`
NotebookLM / infographic / data-viz styles	`references/slide-deck-styles.md`
Seed & temperature (Gemini only)	`references/temperature-seed.md`
Errors, debugging, RONE limitations	`references/troubleshooting.md`
Temperature & seed testing data	`../EXPERIMENT_RESULTS.md`

nano-banana