Intermediate

Generate Scroll-Stopping Thumbnails in 30 Seconds

What this gives you: Type one command in Claude Code, get 3 professional YouTube thumbnails using your real photos. No Photoshop. No Canva. Just a prompt and your headshot. About $0.24 per batch of 3.

Why This Works

Traditional thumbnails need a designer or hours in Canva. With Claude Code and fal.ai's Nano Banana 2 model, you describe what you want and get it in seconds. The key is using your real photos as the base so your face is preserved perfectly. The model transforms the lighting, background, and text - but leaves you looking exactly like you.

This is not a filter. It is a full cinematic transformation built on top of a real photo. The result looks like a professional shoot because the raw material is a real photo. AI-generated faces always look slightly off. Real photos never do.

The Stack

Tool	Role	Cost
fal.ai MCP	Image generation via Nano Banana 2	~$0.08 per thumbnail
Your headshot photos	Source material - local folder or Google Drive	Free
Claude Code	Orchestrates the skill, selects photos, runs generation	Your subscription
Trello (optional)	Attach finished thumbnails to video cards	Free tier works

Total cost per batch of 3 thumbnails: approximately $0.24. You can run this every day without it registering on a budget.

Step 1: Connect the fal.ai MCP Server

Open your Claude Code settings file at ~/.claude/settings.json and add the fal.ai MCP server under the mcpServers key.

# ~/.claude/settings.json
{
  "mcpServers": {
    "fal-ai": {
      "command": "npx",
      "args": ["-y", "@fal-ai/mcp"],
      "env": {
        "FAL_KEY": "${FAL_KEY}"
      }
    }
  }
}

Get your API key from the fal.ai dashboard under API Keys. Then add it to your ~/.env file - never hardcode it in settings.

# ~/.env
FAL_KEY=your-fal-api-key-here

Restart Claude Code after saving. You should see the fal-ai tools appear when you type /mcp in a fresh session.

Step 2: Prepare Your Source Photos

Take 5-6 photos of yourself against a neutral background - a plain wall or standing in a well-lit room. You need variety in expression, not variety in setting. Shoot these in one session and you will never need to reshoot.

Expression	Use for
Shocked - hands on head	Controversial claims, surprising results, "I tried X for 30 days"
Confident - arms crossed	Authority takes, "why most people are wrong about X"
Pointing at camera	Direct address, "you need to see this", tutorials
Calm - slight smile	Step-by-step guides, beginner content, reassuring topics
Leaning in - intense	Secrets, frameworks, premium content reveals

Save all photos to a local folder: ~/headshots/. These are your permanent source files. Name them by expression: shocked.jpg, confident.jpg, pointing.jpg, etc. so the skill can reference them by name.

Upload each photo to fal.media once using the MCP upload tool. Save the returned URLs in a reference file alongside your skill - you will reuse them across every thumbnail batch.

Step 3: Create the Skill File

Create a new skill directory and write the SKILL.md definition.

mkdir -p ~/.claude/skills/youtube-thumbnail

Open ~/.claude/skills/youtube-thumbnail/SKILL.md and write the following:

---
name: youtube-thumbnail
description: >
  Generate 3 YouTube thumbnail variations from a concept direction.
  Uses fal.ai Nano Banana 2 with real headshot photos. Optionally
  attaches to a Trello card. Use when the user needs thumbnails for
  a video or mentions thumbnail generation.
triggers:
  - "thumbnail"
  - "generate thumbnails"
  - "youtube thumbnail"
  - "make me a thumbnail"
---

# /youtube-thumbnail Skill

Generate 3 professional YouTube thumbnails using real headshot photos
and fal.ai Nano Banana 2. Takes a concept direction and optional Trello
card ID.

## Usage
`/youtube-thumbnail [trello-card-id] [concept]`

Example: `/youtube-thumbnail abc123 "I automated my entire business with AI"`

## Steps

### 1. Parse inputs
- Extract the Trello card ID (if provided)
- Extract the video concept/title
- If concept is missing, ask: "What is the video about?"

### 2. Select source photo
- Read the headshot reference file at `~/.claude/skills/youtube-thumbnail/headshots.md`
- Based on the concept tone, select the best expression:
  - Controversial or surprising: shocked
  - Authority or contrarian: confident
  - Tutorial or direct: pointing
  - Calm or beginner-friendly: calm
  - Reveal or insider knowledge: intense

### 3. Construct 3 prompt variations
- Variation 1: Cinematic dark background, bold white text
- Variation 2: Split scene with strong contrast lighting
- Variation 3: Movie poster style with dramatic rim light
- Use the winning prompt formula from the rules below

### 4. Generate thumbnails via fal.ai
- Model: `fal-ai/nano-banana-2/edit`
- Strength: 0.52 (preserves face, allows transformation)
- Aspect ratio: 16:9
- Generate all 3 in parallel

### 5. Attach to Trello (if card ID provided)
- Use the Trello MCP to attach each image URL to the card
- Add a comment: "Thumbnails generated - 3 variations attached"

### 6. Present results
- Show each thumbnail URL with its variation label
- Ask: "Which variation do you prefer?"

## Rules
- ALWAYS use real headshot photos. Never generate a face from text.
- NEVER use Kling models - they mangle faces and text.
- Strength must stay between 0.50 and 0.55. Higher breaks the face.
- Maximum 3 elements per thumbnail: face + text + one prop or icon.
- White text only. No neon colours.
- Glow effects on the wall behind the person, never on skin.
- Dark moody backgrounds only - no white or bright backgrounds.
- Orange/terracotta for Claude branding content, not purple.
- NEVER use adjectives like "big", "bold", or "enormous" in prompts.
  The model renders them as literal text in the image.

## Output
Present the 3 thumbnail URLs with variation labels. Ask which to keep.
If Trello card was provided, confirm attachment.

Also create a headshots.md file in the same directory that maps expression names to fal.media URLs. Update this file once after uploading your photos.

# ~/.claude/skills/youtube-thumbnail/headshots.md

## Headshot URLs

shocked: https://fal.media/files/your-shocked-photo-url.jpg
confident: https://fal.media/files/your-confident-photo-url.jpg
pointing: https://fal.media/files/your-pointing-photo-url.jpg
calm: https://fal.media/files/your-calm-photo-url.jpg
intense: https://fal.media/files/your-intense-photo-url.jpg

The Winning Prompt Formula

This is the exact prompt structure that consistently produces cinematic results. Copy it, fill in the brackets, and feed it to Nano Banana 2 via the edit endpoint.

Transform this into a cinematic YouTube thumbnail. Keep the person
exactly as they are - same face, same [EXPRESSION] expression, same
clothing, same skin tones. Position the person large on the right side
taking up about 40 percent of the frame, cropped from chest up.

Professional studio lighting on the person - soft key light from the
front-left, warm [COLOR] rim light on the right edge. The background
should be a rich dark-to-warm gradient. Add a glow on the wall behind
the person, not on the skin.

At the top, add white text saying "[TEXT]". Keep the text clean and
readable. Deep cinematic shadows, 16:9 aspect ratio.

Fill in: [EXPRESSION] matches your photo, [COLOR] is usually orange or amber for warmth, and [TEXT] is 3-6 words maximum. Short text renders cleanly. Long text breaks.

Style Rules That Actually Matter

These come from testing hundreds of thumbnails. Each rule exists because someone broke it and the result was unusable.

Real photos only. AI-generated faces fail the uncanny valley test. Your subscribers recognise you. Use your actual face.
Nano Banana 2 edit mode only. The endpoint is fal-ai/nano-banana-2/edit. The non-edit endpoint does not preserve identity.
Strength 0.50 to 0.55. Below 0.50, the transformation is too subtle. Above 0.55, the face starts to drift. 0.52 is the sweet spot.
Max 3 elements. Face, text, one additional prop or icon. More than 3 elements create visual noise that kills click-through.
White text only. Neon green, yellow, red - all perform worse on dark backgrounds. White pops on any background.
Glow on the wall, not the skin. Skin glow looks like a filter. Wall glow looks like a film set.
Dark backgrounds always. Bright white backgrounds shrink the thumbnail in the feed. Dark backgrounds make it expand visually.
Orange for Claude branding. If your content is about AI or Claude Code, use orange/terracotta accents. Purple is associated with competitors.

5 Template Prompts

Copy any of these directly. Replace the bracketed sections with your specifics.

1. Cinematic Shocked

Transform this into a cinematic YouTube thumbnail. Keep the person
exactly as they are - same face, same shocked expression, same clothing,
same skin tones. Position them large on the right side, cropped chest up,
taking up 40 percent of the frame. Dark charcoal-to-black gradient
background with a warm orange glow on the wall behind them. Soft studio
key light from front-left, amber rim light on the right edge. White text
at the top: "[YOUR HOOK TEXT]". Deep cinematic shadows. 16:9 ratio.

2. Authority Confident

Transform this into a cinematic YouTube thumbnail. Preserve the person
exactly - same face, confident arms-crossed posture, same clothing.
Position them centre-left, cropped from waist up. Rich deep navy
background with cool-toned rim lighting and a subtle teal glow on the
wall. Professional key light from front-right. White sans-serif text
on the right side: "[YOUR CLAIM]". Sharp, authoritative look. 16:9.

3. Split Contrast Before/After

Transform this into a cinematic YouTube thumbnail. Keep the person
exactly as they are. Position them on the right half of a vertically
split frame. Left half: desaturated grey with gloomy flat lighting.
Right half (where they stand): warm amber glow, cinematic lighting,
rich dark background. A faint dividing line separates the two sides.
White text at the top spanning both halves: "[BEFORE vs AFTER]". 16:9.

4. Contradictory with Strikethrough

Transform this into a cinematic YouTube thumbnail. Preserve the person
exactly - same face, same calm expression, same clothing. Position them
right-of-centre, cropped from chest up. Deep burgundy-to-black gradient
background. Dramatic split lighting - cold blue on one side, warm amber
on the other. Two lines of white text at the top: first line has a
red strikethrough effect over it: "[WRONG BELIEF]". Second line below
in clean white: "[CORRECT BELIEF]". 16:9 aspect ratio.

5. Movie Poster

Transform this into a cinematic YouTube thumbnail styled like a movie
poster. Keep the person exactly as they are - same face, intense
expression, same clothing. Position them centre-frame, full-height from
head to mid-torso. Very dark atmospheric background - near-black with
subtle deep teal fog effect. Strong rim lighting from above creating
a dramatic silhouette edge. Single line of white text below them in
a clean font: "[TITLE]". Cinematic depth of field. 16:9.

The Full Workflow

Here is what happens end-to-end when you invoke the skill with a Trello card ID:

# You type this in Claude Code
/youtube-thumbnail abc123xyz "I replaced my entire team with AI agents"

# Claude:
Reads the Trello card abc123xyz for video context
Opens ~/.claude/skills/youtube-thumbnail/headshots.md
Selects "shocked" expression (matches the concept tone)
Constructs 3 prompt variations using the winning formula
Calls fal-ai/nano-banana-2/edit with strength 0.52 - 3x in parallel
Gets back 3 image URLs (takes ~15 seconds total)
Attaches all 3 to Trello card abc123xyz
Adds comment: "Thumbnails generated - 3 variations attached"
Shows you the URLs and asks which variation to keep

Tip: Generate 3 variations per video and A/B test them. The cost difference is negligible and you will find patterns in what works for your audience. After 10 videos, you will know exactly which expression and lighting style drives the most clicks on your channel.

What NOT to Do

These are the mistakes that consistently produce unusable thumbnails. Avoid every one of them.

Mistake	Result	Fix
Using Kling models	Text in image gets mangled, faces distort	Use Nano Banana 2 edit mode only
AI-generating your face from text	Uncanny valley, subscribers do not recognise you	Always use a real headshot as source
Strength above 0.55	Face drifts, skin tone shifts, identity lost	Stay at 0.50 to 0.55 - 0.52 is optimal
More than 3 visual elements	Visual noise, lower click-through rate	Face, text, and one prop maximum
Neon or coloured text	Hard to read, looks cheap on dark backgrounds	White text only
Outline or cutout effects on the person	Looks like a 2010 YouTube thumbnail	Keep the person naturally integrated into the scene
Adjectives like "big" or "enormous" in the prompt	The model renders those words as text in the image	Describe size with layout instructions, not adjectives

Warning: Never use adjectives like "big", "bold", or "enormous" in your prompts. The model interprets them as text to render and they show up visibly in the generated image. Describe what you want through positioning and lighting, not size adjectives.

Next Guide

AI Ad Creative Generator

Generate Scroll-Stopping Thumbnails in 30 Seconds

Why This Works

The Stack

Step 1: Connect the fal.ai MCP Server

Step 2: Prepare Your Source Photos

Step 3: Create the Skill File

The Winning Prompt Formula

Style Rules That Actually Matter

5 Template Prompts

1. Cinematic Shocked

2. Authority Confident

3. Split Contrast Before/After

4. Contradictory with Strikethrough

5. Movie Poster

The Full Workflow

What NOT to Do

Want a thumbnail system built for your channel?