YouTube Thumbnail A/B Pack

A video topic → 3 distinct 16:9 thumbnail variants to A/B test.

Category: social · Difficulty: Intermediate · ~Time: 9 min

What you'll make

Three high-CTR 16:9 (1280×720) YouTube thumbnails for the same video, each taking a different angle — curiosity, emotion, and bold-claim — so you can split-test which drives the most clicks. Use this before publishing, or to refresh an underperforming video.

Models used

💬 LLM — Claude (generate 3 distinct thumbnail concepts + the 2–4 word overlay text for each)
🎨 Image — Nano Banana Pro (resolve live; e.g. google/nano-banana-pro/text-to-image) and Flux 2 (resolve live; e.g. flux2/flex)

Steps

Generate 3 angles — "For a video titled '', give me 3 thumbnail concepts (curiosity / emotion / bold-claim), each with a 2–4 word overlay and a visual direction." → <code>atlas_chat</code> with <strong>Claude</strong>.</li> <li><strong>Render variant A & B</strong> — "16:9 YouTube thumbnail, shocked face left, huge text '<overlay>' right, saturated colors, thick outline." → <code>atlas_generate_image</code> with <strong>Nano Banana Pro</strong>.<blockquote> <p>Tip: ask the agent to <code>atlas_search_docs</code> for the model first to confirm the live 16:9 / size parameter before rendering.</p> </blockquote> </li> <li><strong>Render variant C</strong> — run the third concept through <strong>Flux 2</strong> for a different rendering aesthetic, so your A/B set isn't visually samey.</li> <li><strong>Check legibility</strong> — view each at small size; the text must read on a phone. Regenerate any that look cluttered.</li> <li><strong>Export the pack</strong> — three 1280×720 PNGs, labeled A/B/C for your test.</li> </ol> <h2>Prompt starters</h2> <pre><code>Concepts (Claude): "Video: '[TITLE]'. Give 3 thumbnail concepts — (1) curiosity gap, (2) emotional reaction, (3) bold claim. For each: a 2–4 word overlay + one-sentence visual." Variant (Nano Banana Pro): "16:9 YouTube thumbnail, close-up shocked expression on the left third, bold yellow text 'I WAS WRONG' on the right, high saturation, strong rim light, thick black text outline." Variant (Flux 2): "16:9 thumbnail, dramatic before/after split composition, punchy contrast, large minimal caption 'BEFORE → AFTER', cinematic lighting." </code></pre> <h2>Tips & variations</h2> <ul> <li><strong>Nano Banana Pro</strong> and <strong>GPT Image 2</strong> are strongest for legible overlay text; <strong>Flux 2</strong> gives a richer photographic look — mixing them keeps the A/B variants genuinely different.</li> <li>Always render true 16:9 (1280×720) so YouTube doesn't crop your text.</li> <li>Push contrast and keep overlay text to ≤4 words — small-size legibility wins clicks.</li> <li>Batch a 4th "control" variant cheaply and let real CTR data pick the winner.</li> </ul> <h2>Related</h2> <ul> <li>More recipes in <a href="../README.html">the library »</a></li> </ul></main> <footer>AI 视频厂商 CLI / Skill / MCP 全景对比 · 技能库 · <a href="/" style="color:var(--muted)">返回报告</a></footer> <script>function t(){const r=document.documentElement;const d=r.dataset.theme==='dark'||(!r.dataset.theme&&matchMedia('(prefers-color-scheme: dark)').matches);r.dataset.theme=d?'light':'dark';}</script></body></html>