generate-image

name	generate-image
description	Generate an image using Google Nano Banana Pro (Imagen). For blog thumbnails, illustrations, diagrams, infographics. Adapted to the Rolebase brand (purple
user_invocable	true
arguments	[{"name":"prompt","description":"Description of the image to generate","required":true},{"name":"output","description":"Output file path relative to website/ (e.g. src/content/blog/my-post/thumbnail.jpg)","required":true},{"name":"size","description":"Image size WxH (default: 1920x1080)","required":false}]

Generate Image

Generate an image using Google Nano Banana Pro (Gemini image generation), styled for the Rolebase brand. This skill is the source of truth for all image generation on the project — other skills (write-blog-post, post-alternative) reference it instead of duplicating these conventions.

The script auto-injects the Rolebase brand into every prompt (warm cream background #FDF6EA, primary purple #9870F0, warm orange #FB9803, warm gray neutrals, friendly human-centric mood). You do not need to restate brand colors in your prompt.

Three modes:

• Illustration (default) — no text, conceptual editorial style. Thumbnails, hero visuals, metaphors.
• Infographic (--infographic) — renders text, numbers, labels. Breakdowns, comparisons, numbered steps, decision matrices.
• Avatar (--avatar) — a uniform flat-vector head-and-shoulders portrait on a plain background. For member/people avatars (e.g. the demo org chart) where a whole set must match.

Usage

Always run from the website/ folder:

cd website

# Illustration (default — no text)
npx tsx scripts/generate-image.ts "<prompt>" <output-path> [WxH]

# Infographic (text and numbers rendered as written in the prompt)
npx tsx scripts/generate-image.ts "<prompt>" <output-path> [WxH] --infographic

# Avatar (uniform flat-vector portrait; defaults to a 512x512 square)
npx tsx scripts/generate-image.ts "<person description>" <output-path> --avatar

Examples:

# Blog thumbnail (1200x630, conceptual, no text)
npx tsx scripts/generate-image.ts "A modern isometric scene of two teams handing off..." src/content/blog/my-post/thumbnail.jpg 1200x630

# In-article infographic (labels and numbers visible)
npx tsx scripts/generate-image.ts "Four numbered cards showing handover steps..." src/content/blog/my-post/steps.jpg 1600x900 --infographic

# Square format
npx tsx scripts/generate-image.ts "An icon..." src/assets/icon.png 1080x1080

# Avatar (only describe the person — style, framing and plain background are enforced by the mode)
npx tsx scripts/generate-image.ts "a woman with shoulder-length wavy auburn hair and light skin, gentle smile" public/demo-avatars/alice.png --avatar

Avatar mode

--avatar produces a flat 2D vector head-and-shoulders portrait with a strict, self-contained style so a whole set of avatars matches: same art style (flat vector, not realistic/3D), same framing and zoom (shoulders and upper chest visible, shoulders at the bottom edge, head not oversized), and a plain flat background (no circle, frame, scene or props — the UI clips avatars to a circle itself). It deliberately skips the editorial brand guidelines, which add scene/context and break uniformity.

• Only describe the person in the prompt (hair, skin, facial hair, glasses, simple top). Do not restate style, framing or background — the mode enforces them.
• The model outputs ~1024px PNG regardless of the requested size; downscale and compress afterwards for web use, e.g. sips -s format jpeg -s formatOptions 80 -Z 192 in.png --out out.jpg (≈8 KB each).
• Generate every avatar of a set with identical wording (vary only the person) so they stay consistent.

Picking the mode

Use infographic when…	Use illustration when…
Numbers, %, ranks need to be visible	The concept is metaphorical or abstract
A comparison table would be more impactful as a visual	The point is mood, not data
Decomposition (steps, components, phases)	A simple table already does the data job
Decision matrix with labels	The visual must be readable at 300px (e.g. thumbnail)

Default to infographics for cost/feature comparisons, breakdowns, and process steps in articles — text in images becomes a featured snippet candidate and LinkedIn previews surface the numbers.

Prompt writing tips

• Be specific about composition, perspective, and spatial layout
• Mention "isometric" or "bird's eye view" for clean staged scenes
• Describe contrast or duality if the article compares two approaches
• Lean into Rolebase themes: roles, org chart, governance, self-management, collaboration, peer feedback. Avoid corporate-pyramid imagery; for people, prefer diverse, casual, modern professionals over stock-suited executives
• Illustration mode: no text (the script forbids it explicitly)
• Infographic mode: write EXACTLY the labels and numbers to render — Nano Banana Pro reproduces text as-is, French diacritics included. Specify the typeface (sans-serif), text color, and position
• Accents (IMPORTANT): always write French text in the prompt WITH its correct accents (é, è, ê, à, ç…). The model renders what you write — if the prompt says "etapes"/"Developpement", the image comes out unaccented and looks wrong. Use the temp-file method (it handles accents fine), never strip diacritics to dodge shell quoting. State explicitly "rendered exactly with correct French accents" and, for titles, "modern clean bold sans-serif" — the model otherwise defaults to an ugly serif for headings.

Blog thumbnail conventions (1200x630)

Thumbnails appear in the blog grid at ~300px wide and as Open Graph previews. They must stand out and stay readable at small sizes. The Rolebase visual language for thumbnails (on top of the brand auto-injection):

• Light, warm cream background — never dark. The brand reads warm and human, not corporate or moody.
• Purple + orange as focal accents — most objects/figures in warm gray, with purple (#9870F0) and orange (#FB9803) glowing on the key element to anchor the eye.
• Soft, warm directional lighting — gentle bloom, smooth gradients. No harsh contrast, no neon.
• Conceptual, not literal — metaphor for the subject (handover, alignment, governance, energy, roles) rather than a direct depiction.
• Human-centric when possible — soft 3D human figures in collaborative postures. Abstract shapes stay organic (rounded, soft, slightly chunky) rather than sharp/technical.
• Single clear focal point — one scene with a strong silhouette, readable at 300px.
• 3D isometric or staged scene — soft, slightly stylized, friendly. Not hyper-realistic, not gamey, not flat illustration.
• No text, no UI elements, no watermarks.

After generation, show the thumbnail to the user. If it lands flat, washed-out, too dark, or too corporate, regenerate with more warmth and a clearer focal accent.

Regeneration comment in MDX articles (image-coupling)

Every generated image placed in an MDX article (infographic or illustration) gets an MDX comment immediately above it, so a future editor can regenerate it without guessing. The comment MUST embed the exact prompt and command used — without the verbatim prompt, a re-run produces a completely different image. For data-driven infographics, also record the regeneration trigger and the last data sync date.

{/* IMAGE regen — npx tsx scripts/generate-image.ts "<exact prompt used, verbatim>" ./image-name.jpg 1600x900 [--infographic]
    Infographic only: regenerate if <data trigger>. Last data sync: YYYY-MM-DD. */}
![alt text](./image-name.jpg)

Keep the prompt in the comment identical to what was passed to the script (accents and apostrophes included). This makes both the regeneration command and the data-image coupling self-documenting.

Shell quoting (IMPORTANT)

Prompts containing apostrophes or French accents (', é, è, à, €...) break inline shell quoting. Workaround:

# Write the prompt to a temp file first
cat > /tmp/prompt.txt <<'EOF'
Your prompt with d'apostrophes and accénts here
EOF

# Then call the script with the contents
PROMPT="$(cat /tmp/prompt.txt)" && npx tsx scripts/generate-image.ts "$PROMPT" src/content/blog/my-post/image.jpg 1600x900 --infographic

# Clean up
rm /tmp/prompt.txt

Requirements

• GOOGLE_AI_API_KEY must be set in website/.env or in the monorepo root .env (the script checks both, in that order)