youtube-thumbnail

YouTube Thumbnail Generation Skill (2026 Edition)

Overview

This skill enables fully autonomous generation of professional YouTube thumbnails in 11 strategic styles from THUMBNAILS.md. Zero user interaction required after the initial request. The agent auto-selects style, builds a prompt, generates the base image via the best available AI backend, applies compositing and text via Pillow, and saves a final 1280×720 PNG to /mnt/user-data/outputs/thumbnail.png.

---

When to Trigger This Skill

Trigger when the user:

• Says "create a thumbnail", "make a thumbnail", "generate a YouTube thumbnail"
• Provides a video title or topic and needs a visual
• Wants product video covers, presentation title slides, or channel art
• Needs batch thumbnail creation or automation for a YouTube workflow

---

Architecture: Two-Layer Pipeline

[User Request]
      │
      ▼
[Layer 1: AI Image Generation]   ← Base image via best available backend
      │
      ▼
[Layer 2: Pillow Compositing]    ← Resize to 1280×720, effects, text overlay
      │
      ▼
[Output: /mnt/user-data/outputs/thumbnail.png]

---

Backend Priority Table

Auto-detect each backend by testing availability. Use the first one that works.

#	Backend	Detection Method	Quality	Cost
1	A1111 Local SD	GET localhost:7860/sdapi/v1/samplers	★★★★	Free (own GPU)
2	ComfyUI Local	GET localhost:8188/history	★★★★	Free (own GPU)
3	MCP Imagen 4 (Vertex AI)	which mcp-imagen-go + ~/.gemini/settings.json	★★★★★	Vertex AI pricing
4	Gemini API (Nano Banana 2)	GEMINI_API_KEY env + google-genai installed	★★★★	Free quota
5	fal.ai FLUX	FAL_KEY env + fal-client installed	★★★★	$0.03/MP
6	OpenAI gpt-image-1	OPENAI_API_KEY env + openai installed	★★★★	~$0.04/img
7	Pillow-only fallback	Always available	★★	Free

---

The 11 Thumbnail Styles

Style 1: Neo-Minimalism (style_1_minimalism)

Best for: General niches, standing out in a cluttered feed Core idea: If the feed is loud, go quiet. 50%+ negative space. AI prompt pattern: "[subject], minimalist product photography, pure white background, single centered subject, dramatic soft studio lighting, ultra clean composition, no clutter" Pillow: White/monochromatic bg, max 2 colors, light serif font bottom-left or none

Style 2: The Surround (style_2_surround)

Best for: Comparisons, "I tried X things", hauls Core idea: Subject dead center, objects in organized circle/grid around it. AI prompt pattern: "[subject] perfectly centered, multiple [related objects] arranged in organized circle or grid around center subject, controlled chaos, vibrant, top-down angle" Pillow: Grid math — center subject 40% canvas, surrounding items equally spaced by angle

Style 3: Rainbow Ranking (style_3_rainbow)

Best for: Tier lists, "Best to Worst", reviews Core idea: Color gradient (Red→Blue) conveys hierarchy visually. AI prompt pattern: "flat lay of [3-7 items] arranged in ranking order, color gradient from red to blue across items, product photography style, clean background" Pillow: Apply gradient color wash per item via ImageEnhance.Color, add rank numbers (1,2,3…) in bold white

Style 4: Educational Whiteboard (style_4_whiteboard)

Best for: Tutorials, business explainers, complex systems Core idea: Authenticity over polish. Signals "high value, no fluff." AI prompt pattern: "hand-drawn diagram on real whiteboard explaining [concept], chalk markers, rough sketchy educational style, authentic classroom feel, [topic] framework" Pillow: Reduce saturation to 70% for authenticity, warm color grade, handwritten-style font

Style 5: Familiar Interface (style_5_ui_framing)

Best for: Commentary, news, reviews Core idea: Borrow credibility from known platforms (Twitter, Reddit, Amazon). AI prompt pattern: "realistic screenshot mockup of [Twitter post / Reddit thread / Amazon listing / Netflix menu] about [topic], exact platform UI styling, authentic spacing and fonts" Pillow: Programmatically draw platform UI elements — rounded rectangles, brand colors (Twitter #1DA1F2, Reddit #FF4500, Amazon #FF9900)

Style 6: Cinematic Text (style_6_cinematic)

Best for: High-production storytelling, documentaries Core idea: Text IS a design element — embedded in the world, not floating over it. AI prompt pattern: "cinematic movie still about [subject], dramatic chiaroscuro lighting, film grain, anamorphic lens flare, shallow depth of field, golden hour or moody tones" Pillow: MAX 3-4 words, large centered bold font, text shadow/glow via layered offset draws

Style 7: Warped Faces (style_7_warped)

Best for: Self-improvement, "Harsh Truths", psychology topics Core idea: "Something is wrong" curiosity gap via distortion. AI prompt pattern: "double exposure portrait, digital glitch effect, [emotion] face merged with [abstract concept], surreal digital distortion, moody dark tones, experimental photography" Pillow: RGB channel shift for glitch (shift R channel +8px), selective blur, minimal/no text

Style 8: Maximalist Flex (style_8_maximalist)

Best for: Collectors, tech enthusiasts, hobbyists Core idea: The collection is the star, not the person. AI prompt pattern: "aerial flat lay of complete collection of every [item type], perfectly organized and arranged, every single item visible, product catalog photography style" Pillow: Dense but organized placement, optional "COMPLETE COLLECTION" text strip top/bottom

Style 9: Encyclopedia Grid (style_9_encyclopedia)

Best for: "Every X Explained", deep dives Core idea: Looks informative and "safe" — no drama, just knowledge. AI prompt pattern: "flat icon illustration grid of [topic] elements, consistent icon shapes, high contrast on white background, educational infographic style, no dramatic lighting" Pillow: Draw equal grid cells with ImageDraw.rectangle, flat icon each cell, label below

Style 10: Candid Fake (style_10_candid)

Best for: Challenges, travel, lifestyle Core idea: Highly engineered to look like a lucky candid shot. AI prompt pattern: "candid authentic moment of [person/scene], natural spontaneous composition but perfectly framed, golden hour lighting, documentary photography style, physically possible scene" Pillow: Minimal processing, subtle vignette at edges only. NO text. NO arrows.

Style 11: The Anti-Thumbnail (style_11_anti)

Best for: Productivity, "Quick Tip" videos Core idea: Dark + specific "irritating" number triggers curiosity. AI prompt pattern: "dark moody portrait of [subject], direct serious eye contact to camera, dramatic low-key lighting, minimal background, cinematic single-subject composition" Pillow: Dark gradient bg (0,0,0)→(30,30,30), specific non-round number ("47 Seconds" not "60"), large centered font

---

Auto-Style Selection (when user doesn't specify)

NICHE_TO_STYLE = {
    # Education & Learning
    "education": "style_4_whiteboard",
    "tutorial": "style_4_whiteboard",
    "howto": "style_4_whiteboard",
    "explainer": "style_9_encyclopedia",
    "course": "style_4_whiteboard",

    # Reviews & Rankings
    "review": "style_3_rainbow",
    "comparison": "style_2_surround",
    "tierlist": "style_3_rainbow",
    "ranking": "style_3_rainbow",
    "top10": "style_3_rainbow",

    # News & Commentary
    "news": "style_5_ui_framing",
    "commentary": "style_5_ui_framing",
    "reaction": "style_5_ui_framing",
    "opinion": "style_5_ui_framing",

    # Personal Development
    "productivity": "style_11_anti",
    "psychology": "style_7_warped",
    "selfimprovement": "style_7_warped",
    "motivation": "style_11_anti",

    # Collections & Gear
    "collection": "style_8_maximalist",
    "tech": "style_8_maximalist",
    "gear": "style_8_maximalist",
    "unboxing": "style_2_surround",

    # Lifestyle & Travel
    "travel": "style_10_candid",
    "lifestyle": "style_10_candid",
    "vlog": "style_10_candid",
    "challenge": "style_10_candid",

    # High-Production
    "documentary": "style_6_cinematic",
    "storytelling": "style_6_cinematic",
    "cinematic": "style_6_cinematic",

    # Default
    "general": "style_1_minimalism",
}

def select_style(niche: str, style_override: str = None) -> str:
    if style_override:
        return style_override
    niche_clean = niche.lower().replace(" ", "").replace("-", "")
    for key in NICHE_TO_STYLE:
        if key in niche_clean or niche_clean in key:
            return NICHE_TO_STYLE[key]
    return "style_1_minimalism"

---

Full Python Implementation

When creating a thumbnail, write this complete script to /home/claude/generate_thumbnail.py, then execute it with python3 generate_thumbnail.py. All values in CAPS are filled in by the agent before writing the script.

#!/usr/bin/env python3
"""
YouTube Thumbnail Generator — Auto-generated by Agent
Video: VIDEO_TITLE_PLACEHOLDER
Style: STYLE_PLACEHOLDER
Backend: auto-detected
"""

import os
import sys
import json
import base64
import io
import subprocess
import requests
from PIL import Image, ImageDraw, ImageFont, ImageFilter, ImageEnhance

# ═══════════════════════════════════════════════════════════
# CONFIGURATION — Agent fills these in before writing script
# ═══════════════════════════════════════════════════════════
VIDEO_TITLE  = "FILL_VIDEO_TITLE"
VIDEO_NICHE  = "FILL_NICHE"           # e.g. "tutorial", "review", "travel"
STYLE        = "FILL_STYLE"           # e.g. "style_6_cinematic"
TEXT_OVERLAY = "FILL_TEXT"            # max 4 words; empty string = auto from title
AI_PROMPT    = "FILL_AI_PROMPT"       # full prompt built from style template
OUTPUT_PATH  = "/mnt/user-data/outputs/thumbnail.png"
# ═══════════════════════════════════════════════════════════


# ─────────────────────────────────────────────────────────
# BACKEND DETECTION
# ─────────────────────────────────────────────────────────

def detect_mcp_imagen() -> bool:
    """Check if mcp-imagen-go binary is installed and configured in Gemini CLI."""
    try:
        r = subprocess.run(["which", "mcp-imagen-go"],
                           capture_output=True, text=True, timeout=5)
        if r.returncode != 0:
            return False
    except (FileNotFoundError, subprocess.TimeoutExpired):
        return False
    settings_path = os.path.expanduser("~/.gemini/settings.json")
    if not os.path.exists(settings_path):
        return False
    try:
        with open(settings_path) as f:
            settings = json.load(f)
        return "imagen" in settings.get("mcpServers", {})
    except Exception:
        return False


def detect_gemini_api() -> bool:
    """Check if Gemini API key is set and google-genai is installed."""
    if not os.environ.get("GEMINI_API_KEY"):
        return False
    try:
        import google.genai
        return True
    except ImportError:
        return False


def detect_backend() -> str:
    """Auto-detect best available image generation backend."""
    # Priority 1: Local A1111
    try:
        r = requests.get("http://127.0.0.1:7860/sdapi/v1/samplers", timeout=3)
        if r.status_code == 200:
            print("✓ Backend: A1111 Local")
            return "a1111"
    except Exception:
        pass

    # Priority 2: Local ComfyUI
    try:
        r = requests.get("http://127.0.0.1:8188/history", timeout=3)
        if r.status_code == 200:
            print("✓ Backend: ComfyUI Local")
            return "comfyui"
    except Exception:
        pass

    # Priority 3: MCP Imagen 4 (Vertex AI via Gemini CLI)
    if detect_mcp_imagen():
        print("✓ Backend: MCP Imagen 4 (Vertex AI)")
        return "mcp_imagen"

    # Priority 4: Gemini API (Nano Banana 2)
    if detect_gemini_api():
        print("✓ Backend: Gemini API (Nano Banana 2)")
        return "gemini_api"

    # Priority 5: fal.ai FLUX
    if os.environ.get("FAL_KEY"):
        try:
            import fal_client
            print("✓ Backend: fal.ai FLUX")
            return "fal"
        except ImportError:
            pass

    # Priority 6: OpenAI gpt-image-1
    if os.environ.get("OPENAI_API_KEY"):
        try:
            import openai
            print("✓ Backend: OpenAI gpt-image-1")
            return "openai"
        except ImportError:
            pass

    # Priority 7: Pillow-only fallback
    print("⚠ Backend: Pillow-only (no AI available)")
    return "pillow_only"


# ─────────────────────────────────────────────────────────
# IMAGE GENERATION — one function per backend
# ─────────────────────────────────────────────────────────

def gen_a1111(prompt: str) -> Image.Image:
    payload = {
        "prompt": prompt,
        "negative_prompt": "blurry, low quality, text, watermark, ugly, deformed, cropped",
        "width": 1280,
        "height": 720,
        "steps": 25,
        "cfg_scale": 7,
        "sampler_name": "DPM++ 2M Karras",
        "batch_size": 1,
    }
    r = requests.post("http://127.0.0.1:7860/sdapi/v1/txt2img", json=payload, timeout=120)
    r.raise_for_status()
    data = r.json()
    img_bytes = base64.b64decode(data["images"][0])
    return Image.open(io.BytesIO(img_bytes))


def gen_comfyui(prompt: str) -> Image.Image:
    """Simple ComfyUI text-to-image via basic workflow."""
    workflow = {
        "3": {"inputs": {"text": prompt, "clip": ["4", 1]}, "class_type": "CLIPTextEncode"},
        "4": {"inputs": {"ckpt_name": "sd_xl_base_1.0.safetensors"}, "class_type": "CheckpointLoaderSimple"},
        "5": {"inputs": {"text": "blurry, ugly, watermark", "clip": ["4", 1]}, "class_type": "CLIPTextEncode"},
        "6": {"inputs": {"width": 1280, "height": 720, "batch_size": 1}, "class_type": "EmptyLatentImage"},
        "7": {"inputs": {"seed": -1, "steps": 25, "cfg": 7, "sampler_name": "dpmpp_2m",
                          "scheduler": "karras", "denoise": 1,
                          "model": ["4", 0], "positive": ["3", 0],
                          "negative": ["5", 0], "latent_image": ["6", 0]},
              "class_type": "KSampler"},
        "8": {"inputs": {"samples": ["7", 0], "vae": ["4", 2]}, "class_type": "VAEDecode"},
        "9": {"inputs": {"images": ["8", 0], "filename_prefix": "thumb"},
              "class_type": "SaveImage"},
    }
    r = requests.post("http://127.0.0.1:8188/prompt",
                      json={"prompt": workflow}, timeout=120)
    r.raise_for_status()
    prompt_id = r.json()["prompt_id"]

    # Poll for result
    import time
    for _ in range(60):
        time.sleep(2)
        hist = requests.get(f"http://127.0.0.1:8188/history/{prompt_id}", timeout=10).json()
        if prompt_id in hist:
            outputs = hist[prompt_id]["outputs"]
            for node_id, node_output in outputs.items():
                if "images" in node_output:
                    img_info = node_output["images"][0]
                    img_r = requests.get(
                        f"http://127.0.0.1:8188/view?filename={img_info['filename']}"
                        f"&subfolder={img_info.get('subfolder','')}&type={img_info['type']}",
                        timeout=30
                    )
                    return Image.open(io.BytesIO(img_r.content))
    raise TimeoutError("ComfyUI generation timed out after 120s")


def gen_mcp_imagen(prompt: str) -> Image.Image:
    """
    Call mcp-imagen-go directly via STDIO (MCP protocol).
    Requires mcp-imagen-go binary in PATH and PROJECT_ID env var.
    Uses Imagen 4 via Vertex AI — highest quality option.
    """
    os.makedirs("/tmp/thumbnail_gen", exist_ok=True)

    mcp_request = json.dumps({
        "jsonrpc": "2.0",
        "method": "tools/call",
        "id": 1,
        "params": {
            "name": "imagen_t2i",
            "arguments": {
                "prompt": prompt,
                "aspect_ratio": "16:9",
                "number_of_images": 1,
                "output_directory": "/tmp/thumbnail_gen",
            }
        }
    })

    # Load PROJECT_ID from settings.json if not in env
    env = os.environ.copy()
    if not env.get("PROJECT_ID"):
        try:
            settings_path = os.path.expanduser("~/.gemini/settings.json")
            with open(settings_path) as f:
                settings = json.load(f)
            mcp_env = settings.get("mcpServers", {}).get("imagen", {}).get("env", {})
            env.update({k: v for k, v in mcp_env.items() if v and "YOUR_" not in v})
        except Exception:
            pass

    proc = subprocess.run(
        ["mcp-imagen-go"],
        input=mcp_request,
        capture_output=True,
        text=True,
        timeout=90,
        env=env
    )

    if proc.returncode != 0:
        raise RuntimeError(f"mcp-imagen-go failed: {proc.stderr[:500]}")

    try:
        response = json.loads(proc.stdout)
    except json.JSONDecodeError:
        # Some versions output multiple JSON lines — take last valid one
        for line in reversed(proc.stdout.strip().split("\n")):
            try:
                response = json.loads(line)
                break
            except Exception:
                continue
        else:
            raise RuntimeError("Could not parse mcp-imagen-go output")

    content = response.get("result", {}).get("content", [])

    for block in content:
        # Inline base64 image
        if block.get("type") == "image" and block.get("data"):
            img_bytes = base64.b64decode(block["data"])
            return Image.open(io.BytesIO(img_bytes))

        # File path returned as text
        if block.get("type") == "text":
            text = block.get("text", "")
            for token in text.split():
                token = token.strip(".,\"'")
                if token.endswith((".png", ".jpg", ".jpeg")) and os.path.exists(token):
                    return Image.open(token)

    # Check output directory for newly created files
    import glob, time
    recent = sorted(glob.glob("/tmp/thumbnail_gen/*.png"), key=os.path.getmtime, reverse=True)
    if recent:
        return Image.open(recent[0])

    raise ValueError("mcp-imagen-go returned no image data")


def gen_gemini_api(prompt: str) -> Image.Image:
    """
    Generate via Gemini API (Nano Banana 2 / gemini-3.1-flash-image-preview).
    Note: aspect ratio is requested in the prompt, not as a parameter.
    """
    from google import genai
    from google.genai import types

    client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

    full_prompt = (
        f"{prompt}. "
        "Generate as a wide landscape 16:9 format, high resolution, "
        "professional YouTube thumbnail quality."
    )

    # Try models newest-first
    models = [
        "gemini-3.1-flash-image-preview",
        "gemini-2.5-flash-image-preview",
        "gemini-2.0-flash-exp",
    ]

    for model_id in models:
        try:
            response = client.models.generate_content(
                model=model_id,
                contents=[full_prompt],
                config=types.GenerateContentConfig(
                    response_modalities=["IMAGE", "TEXT"]
                )
            )
            for part in response.candidates[0].content.parts:
                if part.inline_data is not None:
                    img = Image.open(io.BytesIO(part.inline_data.data))
                    print(f"  ↳ Used model: {model_id}")
                    return img
        except Exception as e:
            print(f"  ↳ {model_id} failed: {e}")
            continue

    raise ValueError("All Gemini API models failed — check GEMINI_API_KEY and quota")


def gen_fal(prompt: str) -> Image.Image:
    import fal_client

    result = fal_client.subscribe(
        "fal-ai/flux/dev",
        arguments={
            "prompt": prompt,
            "image_size": {"width": 1280, "height": 720},
            "num_inference_steps": 28,
            "num_images": 1,
            "enable_safety_checker": True,
        }
    )
    img_url = result["images"][0]["url"]
    r = requests.get(img_url, timeout=60)
    r.raise_for_status()
    return Image.open(io.BytesIO(r.content))


def gen_openai(prompt: str) -> Image.Image:
    from openai import OpenAI
    client = OpenAI()

    response = client.images.generate(
        model="gpt-image-1",
        prompt=prompt,
        size="1536x1024",   # closest 16:9 available
        quality="standard",
        n=1,
    )
    img_bytes = base64.b64decode(response.data[0].b64_json)
    img = Image.open(io.BytesIO(img_bytes))
    # gpt-image-1 returns 1536×1024 — resize to exact YouTube spec
    return img.resize((1280, 720), Image.LANCZOS)


def gen_pillow_only(style: str, title: str) -> Image.Image:
    """
    Pure Pillow fallback — generates a styled graphic without any AI.
    Produces a usable thumbnail when no AI backend is available.
    """
    canvas = Image.new("RGB", (1280, 720))
    draw = ImageDraw.Draw(canvas)

    # Style-specific color palettes
    PALETTES = {
        "style_1_minimalism":  [(245,245,245), (220,220,220)],
        "style_6_cinematic":   [(8, 12, 25),   (40, 30, 60)],
        "style_11_anti":       [(5, 5, 8),     (20, 20, 30)],
        "style_4_whiteboard":  [(250,248,240),  (230,225,210)],
        "style_7_warped":      [(10, 5, 20),   (50, 10, 60)],
        "default":             [(15, 20, 40),  (40, 60, 100)],
    }
    colors = PALETTES.get(style, PALETTES["default"])

    # Vertical gradient
    for y in range(720):
        t = y / 719
        r_v = int(colors[0][0] * (1 - t) + colors[1][0] * t)
        g_v = int(colors[0][1] * (1 - t) + colors[1][1] * t)
        b_v = int(colors[0][2] * (1 - t) + colors[1][2] * t)
        draw.line([(0, y), (1280, y)], fill=(r_v, g_v, b_v))

    # Decorative diagonal accent lines
    accent = (80, 120, 200) if colors[0][0] < 50 else (150, 150, 160)
    for i in range(0, 1280, 120):
        draw.line([(i, 0), (i + 400, 720)], fill=accent, width=1)

    return canvas


# ─────────────────────────────────────────────────────────
# STYLE EFFECTS (Pillow post-processing)
# ─────────────────────────────────────────────────────────

def apply_style_effects(img: Image.Image, style: str) -> Image.Image:
    """Apply style-specific color grading and effects."""

    if style == "style_1_minimalism":
        img = ImageEnhance.Color(img).enhance(0.75)
        img = ImageEnhance.Brightness(img).enhance(1.05)

    elif style == "style_3_rainbow":
        img = ImageEnhance.Color(img).enhance(1.4)
        img = ImageEnhance.Contrast(img).enhance(1.1)

    elif style == "style_4_whiteboard":
        img = ImageEnhance.Color(img).enhance(0.65)
        # Warm tone shift
        r, g, b = img.split()
        r = ImageEnhance.Brightness(Image.merge("RGB", (r, r, r))).enhance(1.05).split()[0]
        img = Image.merge("RGB", (r, g, b))

    elif style == "style_6_cinematic":
        img = ImageEnhance.Color(img).enhance(0.85)
        img = ImageEnhance.Contrast(img).enhance(1.3)
        # Slight teal shadow / orange highlight look
        img = _apply_color_grade(img, shadow=(0, 5, 15), highlight=(15, 5, 0))

    elif style == "style_7_warped":
        # RGB channel shift for glitch effect
        r, g, b = img.split()
        r = r.transform(img.size, Image.AFFINE, (1, 0, 8, 0, 1, 0))
        b = b.transform(img.size, Image.AFFINE, (1, 0, -6, 0, 1, 2))
        img = Image.merge("RGB", (r, g, b))
        img = ImageEnhance.Contrast(img).enhance(1.2)

    elif style == "style_9_encyclopedia":
        img = ImageEnhance.Color(img).enhance(0.6)
        img = ImageEnhance.Brightness(img).enhance(1.1)

    elif style == "style_11_anti":
        img = ImageEnhance.Brightness(img).enhance(0.6)
        img = ImageEnhance.Contrast(img).enhance(1.5)

    else:
        # Default: moderate contrast boost
        img = ImageEnhance.Contrast(img).enhance(1.15)

    # Vignette applied to all styles
    img = _apply_vignette(img, strength=0.35)

    return img


def _apply_color_grade(img: Image.Image,
                        shadow=(0, 0, 0),
                        highlight=(0, 0, 0)) -> Image.Image:
    """Subtle shadow/highlight color grade (like LUTs)."""
    r, g, b = img.split()

    def grade_channel(channel, shadow_add, highlight_add):
        lut = []
        for i in range(256):
            t = i / 255.0
            val = i + int(shadow_add * (1 - t)) + int(highlight_add * t)
            lut.append(max(0, min(255, val)))
        return channel.point(lut)

    r = grade_channel(r, shadow[0], highlight[0])
    g = grade_channel(g, shadow[1], highlight[1])
    b = grade_channel(b, shadow[2], highlight[2])
    return Image.merge("RGB", (r, g, b))


def _apply_vignette(img: Image.Image, strength: float = 0.35) -> Image.Image:
    """Add subtle radial vignette to focus the eye toward center."""
    w, h = img.size
    mask = Image.new("L", (w, h), 255)
    draw = ImageDraw.Draw(mask)

    steps = min(w, h) // 2
    for i in range(steps):
        progress = i / steps
        alpha = int(255 * (progress + (1 - progress) * (1 - strength)))
        alpha = max(0, min(255, alpha))
        margin_x = int((1 - progress) * (w // 2))
        margin_y = int((1 - progress) * (h // 2))
        draw.ellipse(
            [margin_x, margin_y, w - margin_x, h - margin_y],
            fill=alpha
        )

    mask = mask.filter(ImageFilter.GaussianBlur(radius=40))
    black = Image.new("RGB", (w, h), (0, 0, 0))
    img = Image.composite(black, img, mask)
    return img


# ─────────────────────────────────────────────────────────
# TEXT OVERLAY
# ─────────────────────────────────────────────────────────

FONT_PATHS = [
    "/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf",
    "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf",
    "/usr/share/fonts/TTF/DejaVuSans-Bold.ttf",
    "/usr/share/fonts/truetype/freefont/FreeSansBold.ttf",
    "/usr/share/fonts/truetype/ubuntu/Ubuntu-Bold.ttf",
]

# Styles where text should be omitted
NO_TEXT_STYLES = {"style_7_warped", "style_10_candid"}


def load_font(size: int):
    for fp in FONT_PATHS:
        if os.path.exists(fp):
            try:
                return ImageFont.truetype(fp, size=size)
            except Exception:
                continue
    return ImageFont.load_default()


def add_text_overlay(img: Image.Image, text: str, style: str) -> Image.Image:
    """Add styled text overlay appropriate for each thumbnail style."""

    if not text or style in NO_TEXT_STYLES:
        return img

    # Truncate to 4 words max (per best-practice from THUMBNAILS.md)
    words = text.split()
    if len(words) > 4:
        text = " ".join(words[:4])

    w, h = img.size
    img = img.convert("RGBA")

    if style in ("style_6_cinematic", "style_11_anti"):
        return _text_centered_large(img, text, w, h)

    elif style in ("style_1_minimalism", "style_4_whiteboard"):
        return _text_clean_corner(img, text, w, h, style)

    elif style == "style_11_anti":
        return _text_centered_large(img, text, w, h)

    else:
        return _text_banner_strip(img, text, w, h)


def _text_centered_large(img, text, w, h):
    """Large centered text for cinematic/anti-thumbnail styles."""
    font = load_font(96)
    draw = ImageDraw.Draw(img)

    bbox = draw.textbbox((0, 0), text.upper(), font=font)
    tw, th = bbox[2] - bbox[0], bbox[3] - bbox[1]
    x, y = (w - tw) // 2, (h - th) // 2

    # Glow / shadow effect
    for offset in [(6, 6), (-6, 6), (6, -6), (-6, -6)]:
        draw.text((x + offset[0], y + offset[1]), text.upper(),
                  font=font, fill=(0, 0, 0, 180))
    draw.text((x, y), text.upper(), font=font, fill=(255, 255, 255, 255))

    return img.convert("RGB")


def _text_clean_corner(img, text, w, h, style):
    """Clean minimal text for minimalism and whiteboard styles."""
    font = load_font(72)
    draw = ImageDraw.Draw(img)

    text_color = (30, 30, 30, 255) if style == "style_4_whiteboard" else (60, 60, 60, 255)
    bbox = draw.textbbox((0, 0), text, font=font)
    x, y = 60, h - (bbox[3] - bbox[1]) - 60

    # Subtle shadow
    draw.text((x + 2, y + 2), text, font=font, fill=(200, 200, 200, 120))
    draw.text((x, y), text, font=font, fill=text_color)

    return img.convert("RGB")


def _text_banner_strip(img, text, w, h):
    """Semi-transparent banner strip with high-contrast text."""
    font = load_font(82)
    draw_measure = ImageDraw.Draw(img)
    bbox = draw_measure.textbbox((0, 0), text.upper(), font=font)
    tw, th = bbox[2] - bbox[0], bbox[3] - bbox[1]

    padding_x, padding_y = 30, 18
    strip_h = th + padding_y * 2
    strip_y = h - strip_h - 40

    # Semi-transparent background strip
    overlay = Image.new("RGBA", (w, h), (0, 0, 0, 0))
    overlay_draw = ImageDraw.Draw(overlay)
    overlay_draw.rectangle(
        [0, strip_y, w, strip_y + strip_h],
        fill=(0, 0, 0, 175)
    )
    img = Image.alpha_composite(img, overlay)

    # Text: shadow then main
    draw = ImageDraw.Draw(img)
    x = (w - tw) // 2
    y = strip_y + padding_y

    draw.text((x + 3, y + 3), text.upper(), font=font, fill=(0, 0, 0, 200))
    draw.text((x, y), text.upper(), font=font, fill=(255, 220, 50, 255))

    return img.convert("RGB")


# ─────────────────────────────────────────────────────────
# AUTO TEXT EXTRACTION
# ─────────────────────────────────────────────────────────

def auto_text(video_title: str, style: str) -> str:
    """Extract best text overlay from video title for given style."""
    if style in NO_TEXT_STYLES:
        return ""
    words = video_title.split()
    # For anti-thumbnail: keep number if present, else use 3 words
    if style == "style_11_anti":
        for word in words:
            if any(c.isdigit() for c in word):
                return word + (" Seconds" if "sec" not in video_title.lower() else "")
        return " ".join(words[:3])
    # General: first 4 impactful words
    stopwords = {"the", "a", "an", "how", "to", "i", "my", "is", "are", "was"}
    filtered = [w for w in words if w.lower() not in stopwords]
    result = filtered[:4] if filtered else words[:4]
    return " ".join(result)


# ─────────────────────────────────────────────────────────
# DEPENDENCY INSTALLER
# ─────────────────────────────────────────────────────────

def ensure_deps(backend: str):
    """Install required packages for the selected backend."""
    deps = ["Pillow", "requests"]

    if backend == "gemini_api":
        deps.append("google-genai")
    elif backend == "fal":
        deps.append("fal-client")
    elif backend == "openai":
        deps.append("openai")

    for dep in deps:
        try:
            if dep == "Pillow":
                import PIL
            elif dep == "requests":
                import requests
            elif dep == "google-genai":
                import google.genai
            elif dep == "fal-client":
                import fal_client
            elif dep == "openai":
                import openai
        except ImportError:
            print(f"Installing {dep}...")
            subprocess.run(
                [sys.executable, "-m", "pip", "install", dep,
                 "--break-system-packages", "-q"],
                check=True
            )


# ─────────────────────────────────────────────────────────
# MAIN ORCHESTRATOR
# ─────────────────────────────────────────────────────────

def main():
    print(f"\n🎨 Thumbnail Generator")
    print(f"   Title : {VIDEO_TITLE}")
    print(f"   Style : {STYLE}")
    print(f"   Output: {OUTPUT_PATH}\n")

    # 1. Detect backend
    backend = detect_backend()

    # 2. Install deps if needed
    ensure_deps(backend)

    # 3. Generate base image
    print(f"→ Generating base image...")
    generators = {
        "a1111":       gen_a1111,
        "comfyui":     gen_comfyui,
        "mcp_imagen":  gen_mcp_imagen,
        "gemini_api":  gen_gemini_api,
        "fal":         gen_fal,
        "openai":      gen_openai,
    }

    if backend == "pillow_only":
        base_img = gen_pillow_only(STYLE, VIDEO_TITLE)
    else:
        try:
            base_img = generators[backend](AI_PROMPT)
        except Exception as e:
            print(f"⚠ {backend} failed: {e}")
            print("  Falling back to Pillow-only...")
            base_img = gen_pillow_only(STYLE, VIDEO_TITLE)

    # 4. Normalize to exact 1280×720
    base_img = base_img.convert("RGB").resize((1280, 720), Image.LANCZOS)
    print(f"✓ Base image ready: {base_img.size}")

    # 5. Apply style effects
    print("→ Applying style effects...")
    base_img = apply_style_effects(base_img, STYLE)

    # 6. Determine text overlay
    text = TEXT_OVERLAY if TEXT_OVERLAY else auto_text(VIDEO_TITLE, STYLE)
    print(f"→ Text overlay: '{text}'" if text else "→ No text overlay (style preference)")

    # 7. Add text
    base_img = add_text_overlay(base_img, text, STYLE)

    # 8. Save
    os.makedirs(os.path.dirname(OUTPUT_PATH), exist_ok=True)
    base_img.save(OUTPUT_PATH, "PNG", optimize=True)

    size_kb = os.path.getsize(OUTPUT_PATH) // 1024
    print(f"\n✅ Saved: {OUTPUT_PATH} ({size_kb} KB, 1280×720)")


if __name__ == "__main__":
    main()

---

Agent Execution Protocol

When the user asks for a thumbnail, the agent follows these steps:

Step 1 — Parse

Extract from message:

• VIDEO_TITLE — the video title or topic description
• VIDEO_NICHE — category (tutorial, review, travel, etc.)
• STYLE — if explicitly mentioned; otherwise auto-select
• TEXT_OVERLAY — specific text if mentioned (max 4 words); else leave empty

Step 2 — Select Style

style = select_style(VIDEO_NICHE, style_override=None)

Step 3 — Build AI Prompt

Use the style template from "The 11 Styles" section above. Append universal quality suffix:

", professional YouTube thumbnail, vibrant high contrast, cinematic quality,
sharp focus, award-winning composition"

Step 4 — Fill Script Template

Replace all FILL_ placeholders in the Python script above with actual values.

Step 5 — Execute

pip install Pillow requests --break-system-packages -q
python3 /home/claude/generate_thumbnail.py

Step 6 — Verify & Present

assert os.path.exists("/mnt/user-data/outputs/thumbnail.png")
assert os.path.getsize("/mnt/user-data/outputs/thumbnail.png") > 50_000

Then call present_files tool with the output path. Briefly tell the user which style was chosen and why (1 sentence).

---

Style Prompt Templates (Reference Card)

STYLE_PROMPTS = {
    "style_1_minimalism": (
        "{subject}, minimalist product photography, pure white background, "
        "single centered subject, dramatic soft studio lighting, ultra clean"
    ),
    "style_2_surround": (
        "{subject} dead center, multiple {objects} arranged in perfect organized "
        "circle around center, controlled chaos, vibrant, top-down angle"
    ),
    "style_3_rainbow": (
        "flat lay of {items} in ranking order, color gradient red to blue, "
        "product photography, clean background, vivid colors"
    ),
    "style_4_whiteboard": (
        "hand-drawn diagram on real whiteboard explaining {concept}, chalk markers, "
        "rough sketchy authentic educational style, classroom feel"
    ),
    "style_5_ui_framing": (
        "realistic screenshot mockup of {platform} UI about {topic}, "
        "exact platform styling, authentic spacing and fonts, credible interface"
    ),
    "style_6_cinematic": (
        "cinematic movie still about {subject}, dramatic chiaroscuro lighting, "
        "film grain, anamorphic lens flare, shallow depth of field"
    ),
    "style_7_warped": (
        "double exposure portrait, digital glitch effect, {emotion} face merged "
        "with {concept}, surreal distortion, moody dark tones"
    ),
    "style_8_maximalist": (
        "aerial flat lay of complete collection of all {items}, perfectly organized, "
        "every single item visible, product catalog photography"
    ),
    "style_9_encyclopedia": (
        "flat icon illustration grid of {topic} elements, consistent icon shapes, "
        "high contrast on white background, educational infographic style"
    ),
    "style_10_candid": (
        "candid authentic moment of {scene}, natural spontaneous composition, "
        "perfectly framed, golden hour lighting, documentary photography"
    ),
    "style_11_anti": (
        "dark moody portrait of {subject}, direct serious eye contact to camera, "
        "dramatic low-key lighting, minimal background, cinematic"
    ),
}

---

Output Specification

• Path: /mnt/user-data/outputs/thumbnail.png
• Resolution: 1280 × 720 px (YouTube 16:9 standard)
• Format: PNG
• Min file size: ~50 KB (abort and retry if smaller)

---

Error Handling Rules

1. Backend API fails → automatically fall back to next priority backend
2. All AI backends fail → use gen_pillow_only(), still produce output
3. Font not found → use ImageFont.load_default(), never crash
4. Image < 50KB after save → regenerate with next backend
5. TEXT_OVERLAY blank → run auto_text() to extract from title
6. mcp-imagen-go PATH issue → check ~/.gemini/settings.json env and retry with full binary path from which mcp-imagen-go

---

Quick Reference: Which Backend for What

Situation	Recommended Backend
Local GPU available (A1111 running)	a1111 — always fastest+free
Have GCP project + Gemini CLI set up	mcp_imagen — Imagen 4, best quality
Have Gemini API key (from Gemini CLI)	gemini_api — free quota, good quality
Cloud-only, budget matters	fal — $0.03/image, FLUX quality
Need best text rendering in image	openai — gpt-image-1 best for text
No API keys / testing	pillow_only — always works