# How to make YouTube video thumbnails with AI > **For AI agents:** find the right skill in one call — `curl -s "https://skills.varg.ai/api/search?q="` returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://skills.varg.ai/llms.txt. Every page has a markdown twin: add `.md` to its path. Before generating anything, show the person the options on the how-to page. High-CTR thumbnails and vertical covers, in a few options to choose from. A step-by-step tutorial for Claude Code, Codex and Grok. **Output:** 1280×720 PNGs (and 9:16 covers) **Price:** in credits on varg, with a cost estimate before anything runs; the free skills below install in one command. ## Recommended route - **Start with:** varg AI (`varg-ai`) — On varg, one key: [`gpt_image_2_5`](https://varg.ai/models/gpt_image_2_5.md) or [`nano_banana_2_1`](https://varg.ai/models/nano_banana_2_1.md) for the image, [`ideogram_v4_5`](https://varg.ai/models/ideogram_v4_5.md) when the text must be exact — a few credits per option. - **Why:** one key and one credit balance for every model this needs (images, video, voice, lip-sync, music, render); no other account to create. - **Alternative:** Higgsfield YouTube Thumbnail (`higgsfield-youtube-thumbnail`) — Truthful, high-CTR thumbnails and covers with consistent faces · Needs: Higgsfield account, Higgsfield CLI (`higgsfield auth login`) ## Best skills for this 1. **varg AI** (`varg-ai`, varg, live) — Adds: GPT Image 2.5 / Nano Banana 2.1 generation on one key - Claude Code: `claude plugin marketplace add vargHQ/skills && claude plugin install varg@varg` - Codex: `npx skills add vargHQ/skills --skill varg-ai -a codex -y` - Cursor: `npx skills add vargHQ/skills --skill varg-ai -a cursor -y` - OpenClaw: `npx clawhub@latest install varg-ai` - Details: https://skills.varg.ai/skills/varg-ai.md · Example: https://s3.varg.ai/how-to/previews/7df061df8dcb/varg-ai.mp4 2. **Higgsfield YouTube Thumbnail** (`higgsfield-youtube-thumbnail`, Higgsfield, live) — Start here: Truthful, high-CTR thumbnails and covers with consistent faces · Needs: Higgsfield account, Higgsfield CLI (`higgsfield auth login`) - Claude Code: `/plugin marketplace add higgsfield-ai/skills && /plugin install higgsfield@higgsfield` - Codex: `npx skills add higgsfield-ai/skills --skill higgsfield-youtube-thumbnail -a codex` - Cursor: `npx skills add higgsfield-ai/skills --skill higgsfield-youtube-thumbnail -a cursor` - Gemini CLI: `npx skills add higgsfield-ai/skills --skill higgsfield-youtube-thumbnail -a gemini-cli` - Details: https://skills.varg.ai/skills/higgsfield-youtube-thumbnail.md · Example: https://s3.varg.ai/how-to/previews/ca2be60825df/higgsfield-youtube-thumbnail.webp 3. **YouTube Thumbnail Design** (`inference-sh-youtube-thumbnail-design`, inference.sh, live) — Start here: Thumbnail rules: phone readability, contrast, safe zones, text · Needs: inference.sh account, belt CLI (`belt login`) - Claude Code: `npx skills add inference-sh/skills --skill youtube-thumbnail-design -a claude-code` - Codex: `npx skills add inference-sh/skills --skill youtube-thumbnail-design -a codex` - Cursor: `npx skills add inference-sh/skills --skill youtube-thumbnail-design -a cursor` - Gemini CLI: `npx skills add inference-sh/skills --skill youtube-thumbnail-design -a gemini-cli` - Details: https://skills.varg.ai/skills/inference-sh-youtube-thumbnail-design.md 4. **YouTube MCP** (`youtube-mcp`, Zubeid Hendricks, live) — Adds: Pulls the niche's top thumbnails and view counts for a reference board · Needs: YouTube Data API key (free, 10k units/day) - Any MCP client: `npx -y @smithery/cli install @ZubeidHendricks/youtube --client claude` - Details: https://skills.varg.ai/skills/youtube-mcp.md ## Ask the person first Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it. **Face?** - Your face, a believable emotion — [Higgsfield YouTube Thumbnail](https://skills.varg.ai/skills/higgsfield-youtube-thumbnail.md) · example: https://s3.varg.ai/how-to/previews/ca2be60825df/higgsfield-youtube-thumbnail.webp - No face: one striking object or scene - Several faces (group) **Text?** - No text - 1–3 words ## How to make YouTube video thumbnails: step by step 1. Pull 10–20 top thumbnails from the niche onto a board 2. Write 3 genuinely different concepts (not colour swaps) 3. Generate; check each at 160 px wide 4. Make sure the first 5 seconds of the video deliver the thumbnail's promise 5. Upload all 3 to Test & Compare and keep the watch-time winner ## What works now (researched 2026-10-27) 2026 breakout thumbnails are simpler than the MrBeast-era template. Of 500 breakouts in June 2026, 69% had a face (80% of the top 50), but only ~5% used an exaggerated expression, and no text at all was the single most common choice (28%). YouTube's Test & Compare now picks winners by watch-time share, so a thumbnail that over-promises loses even if it gets the click. 1. **Design for a phone at a glance: one subject, one idea, high contrast. Check it at about 160 px wide before you call it done.** MrBeast: 'on phones thumbnails are really small... When you shrink it down, you can't see anything.' Paddy Galloway: viewers see packaging for 'milliseconds', so the thumbnail must stop the scroll before the title builds curiosity. ([Paddy Galloway, packaging a viral idea (Jul 2026)](https://x.com/PaddyG96/status/2080681431718678741)) 2. **Use no text or 1-3 words. Never repeat the title in the image.** Of 500 breakouts, 28% had no text and 24% had 1-3 words. The median among those with text was 5 words, so long overlays are the minority. ([vidIQ, 500 breakout thumbnails (Jun 2026)](https://vidiq.com/research/youtube-thumbnail-study/)) 3. **Use a real face filling about a third of the frame with a readable, believable emotion when the channel is personality-led. For faceless channels, use one striking object or scene instead.** Faces rise with performance (69% of all breakouts, 80% of the top 50), but exaggerated faces are only ~5%. 1of10's 300k-video study found faces help in some niches (finance), hurt in others (business), and that several faces beat one. ([Search Engine Journal, do faces help thumbnails (1of10 vs vidIQ)](https://searchenginejournal.com/do-faces-help-youtube-thumbnails)) 4. **Make 3 genuinely different concepts, not colour swaps, and run Test & Compare once videos get 5k+ views. Set the winner when watch-time share differs by 7% or more.** Test & Compare picks by watch-time share, not CTR. Rene Ritchie: over-indexing on CTR 'could become click-bait, which could tank retention'. Nate Curtiss finds small-channel tests inconclusive and tests whole angles first. ([Nate Curtiss, how he A/B tests packaging (Jul 2026)](https://x.com/natecurtiss_yt/status/2082493105853374832)) 5. **Show the thumbnail's promise in the first 5 seconds of the video.** Autoplay means many viewers judge the opening, not the thumbnail (MrBeast). Test & Compare scores watch time, and YouTube has started removing 'egregious clickbait' whose title or thumbnail the video doesn't deliver (India first). ([MrBeast on thumbnails, titles and autoplay (clip, Apr 2026)](https://x.com/jaynitx/status/2043657422750040184)) **Stopped working:** - The exaggerated open-mouth shock face as the default: only ~5% of 2026 breakouts use one. - 5+ words of text, or text that repeats the title: it is unreadable at phone size and wastes the second line of curiosity. - Promising something the first 5 seconds don't show. It loses Test & Compare on watch time and risks removal under YouTube's clickbait enforcement. ## Prompts that work Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}. ### Thumbnail render prompt in 11 blocks (nano_banana_2_1) The main render: one prompt per concept with a reference manifest, identity lock, single information-gap element, power-third composition and the thumbnail lighting rig. Adapted from [Higgsfield YouTube Thumbnail skill, prompt contract](https://github.com/higgsfield-ai/skills/blob/main/higgsfield-youtube-thumbnail/SKILL.md). ```text IMAGE REFERENCES: image 1 = CHARACTER 1 face reference{; image 2 = brand logo}. Bold, punchy YouTube-thumbnail composite, poster-grade and photoreal, not a muted movie still, 16:9, one continuous shot, no split-screen, no diagonal divide. SCENE: {THE ONE MOMENT OR OBJECT THE VIDEO IS ABOUT}. TEXT: No text, no readable UI labels, no watermark. {OR: bold headline reading exactly "{1-3 WORDS}", massive outlined sans-serif, never over the face} CHARACTER 1: the person from face reference #1. IDENTITY LOCK: same bone structure, eye shape, nose, lips, jawline, skin tone, hairline and hair texture; do not beautify, average or restyle. Expression: {A BELIEVABLE EMOTION}. Chest-up, foreground, 40-60% of the frame. All faces crisply sharp. KEY ELEMENTS: {ONLY THE PROP OR RESULT THAT OPENS THE INFORMATION GAP}. LOCATION: {PLACE, TIME OF DAY, WEATHER}. COMPOSITION: one hero on a power third, clear scale hierarchy, depth, strong subject-background separation. BACKGROUND: vivid high-contrast {COLOUR} field or environment, soft vignette, undivided. LIGHTING: strong key light sculpting the face, soft fill, defined back light and hair light tracing a clean rim; only the rim may carry a {ACCENT} tint. GRADE: vivid, bright, glossy, deep blacks, crisp highlights, saturated, one cohesive image. ``` ### Niche reference board to 3 concepts (agent) Agent prompt for the whole job: pull the niche's top thumbnails onto a board, borrow structure not identity, render 3 different concepts, check them at phone size. Adapted from [Higgsfield YouTube Thumbnail skill (reference analysis, concept gate, post-render gate)](https://github.com/higgsfield-ai/skills/blob/main/higgsfield-youtube-thumbnail/SKILL.md). ```text Make 3 thumbnail options for my video "{TITLE}". Its truthful promise: {ONE LINE}. The first 5 s show {OPENING SHOT}. 1. Board: pull the top {15} videos in {NICHE} from the last {90} days by views (YouTube Data API or youtube-mcp), download their thumbnails and lay them out on one HTML contact sheet with views and channel size under each. 2. For the 5 strongest, extract only the structure: subject pose or action, elements, location, composition, background, split yes/no, person count, emotion. Never copy a face, a logo or the exact layout. 3. Write 5 truthful concepts across different frameworks (posed portrait, size difference, before/after, repetition of one object, one striking object, map or route). Each opens an information gap the title answers and reads in under a second. Keep 3 that are genuinely different, not colour swaps. 4. Render each with {IMAGE_MODEL} at 1280x720, {FACE_PHOTO} as the identity reference, 0-3 words of text only when they continue the title, never repeat it. 5. Check every result at 160 px wide next to the board: one subject, high contrast, face or hero still readable, no stray text. Retry a failed one at most twice. 6. Show the 3 with a one-line label each, then edit only what I ask (expression, background, rim colour) and keep everything else unchanged. I will upload all 3 to Test & Compare. ``` ## Trending formats to remake To remake one: download the example with the `media-download` skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below. 1. **Face + one object + 1-3 words**: A creator's face at about a third of the frame with a believable reaction, next to the one object or result the video is about, with 0-3 words of text. Why it spreads: It is the most common pattern among top breakouts: faces on 80% of the top 50, most text at 3 words or fewer. Named templates ('I Spent 24 Hours in ___' with an hour counter, 1-star vs 5-star) can be reused across a series. Keep: Your face at about a third of the frame with a believable, not shocked, emotion next to the one object or result the video is about, plus 0-3 words that continue the title; swap the face, the object and the words. Make it with [Higgsfield YouTube Thumbnail](https://skills.varg.ai/skills/higgsfield-youtube-thumbnail.md) (`higgsfield-youtube-thumbnail`). - Example to remake: [@mws on X, 2026-10-23, 25K views, 950 saves](https://x.com/mws/status/2113761935124504579): A talking head on top, but the bottom half scrolls a doc of 234 real viral thumbnails grouped into named formats ('Overnight in the World's Most ___', 'I Spent 24 Hours in ___' with an hour counter, 1-star vs 5-star). Good for picking a template. Save ratio 1.19. 2. **No-text single striking image**: One surprising scene or object, high contrast, no words. The title does all the talking. Why it spreads: No text is the most common single pattern among 500 June-2026 breakouts (28%). It also works for faceless channels and survives YouTube's small mobile sizes. No X example passed the bar. Keep: One surprising scene or object, high contrast, nothing else, still readable at 160 px wide; swap the subject and let the title do the talking. Make it with [YouTube Thumbnail Design](https://skills.varg.ai/skills/inference-sh-youtube-thumbnail-design.md) (`inference-sh-youtube-thumbnail-design`). 3. **Niche-reference remix (3 variants)**: Pull the top 10-20 thumbnails of the niche or channel onto a board, pick the strongest layout, and generate 3-4 variants in that style for your topic. Why it spreads: It turns 'what already wins here' into the starting point and feeds Test & Compare with genuinely different options. The demo was saved 1,119 times. Keep: A board of the niche's top 10-20 thumbnails, borrow the winning layout (never the face or logo), make 3 genuinely different concepts and let Test & Compare pick; swap the topic and the face. Make it with [varg AI](https://skills.varg.ai/skills/varg-ai.md) (`varg-ai`). - Example to remake: [@rileybrown on X, 2026-06-02, 99K views, 1K saves](https://x.com/rileybrown/status/2061634840999448629): Workflow demo, not a single thumbnail: the agent pulls a channel's top 12 thumbnails onto a board, then 'make a thumbnail in the style of the red one except the background should be black' returns 4 variants with GPT Image 2. ## Open-source repos - [higgsfield-ai/skills](https://github.com/higgsfield-ai/skills) (2K★): Agent skill for truthful, high-CTR thumbnails and Shorts covers with up to three consistent faces (runs through the Higgsfield CLI). How: Title + 0-3 face photos + optional logo or style reference -> 5+ truthful concepts across 16 frameworks, best information gap picked -> a style reference is read by vision into structure fields, never passed as an image -> 11-block prompt rendered with Nano Banana Pro at 4K -> vision check at ~120 px -> surgical edits of expression, background or rim only. Install as a skill: https://skills.varg.ai/skills/higgsfield-youtube-thumbnail.md - [inference-sh/skills](https://github.com/inference-sh/skills) (783★): Thumbnail design guide skill: specs, the 120 px test, safe zones, contrast, text and face rules, with generation commands. How: Rules first (1280x720, keep the bottom-right timestamp corner clear, high-contrast pairs, a few words at most, the face on one side looking toward the subject) -> per-content patterns (tutorial, before/after, review) -> example image-model calls through the belt CLI. Install as a skill: https://skills.varg.ai/skills/inference-sh-youtube-thumbnail-design.md - [ZubeidHendricks/youtube-mcp-server](https://github.com/ZubeidHendricks/youtube-mcp-server) (590★): YouTube Data API over MCP: the research step for a reference board. How: videos_searchVideos with order, date and channel-size filters plus video and channel stats -> the agent fetches each top video's thumbnail and grids them with views before writing concepts. Install as a skill: https://skills.varg.ai/skills/youtube-mcp.md ## Set up varg varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs. - Claude Code: `claude plugin marketplace add vargHQ/skills && claude plugin install varg@varg` (skill + hosted MCP) - Codex, Cursor, Gemini CLI, OpenClaw: `npx skills add vargHQ/skills --skill varg-ai -a -y` - Grok and any agent that reads SKILL.md: `npx skills add vargHQ/skills --skill varg-ai`, or point it at the skill's SKILL.md - Claude.ai, ChatGPT, any MCP client: connect `https://mcp.varg.ai/mcp` - API key: https://varg.ai/dashboard/api · Model specs: https://varg.ai/models/llms.txt ## Related tutorials - [How to make TikTok photo carousels with AI](https://skills.varg.ai/make-tiktok-photo-carousels.md): Swipeable photo slideshows with hook text — the format app founders use for organic growth.