The guide · Thumbnails

The cover lock: thumbnails that stay on brand without designing them.

I used to start every cover on a blank canvas. Different font, different layout, different vibe each time. Scroll my grid and it looked like a group project, not one account.

The fix is three rules you set once. After that you stop designing covers. You fill in a prompt.

The three rules

Rule 1 · The type one font, one accent colour. Mine is a clean bold sans with one word in lime. That's the whole system.
Rule 2 · The spot the title sits in the same place on every cover, so your grid reads like a series, not a scrapbook.
Rule 3 · The source never generate a cover from nothing. Give the image model two references: a frame from the actual video and your type sheet. Your face stays your face. Your letters stay your letters.

The prompt, the example and the Claude skill

🔒 Drop your email to reveal the prompts

Get the whole guide, free.

Drop your email to unlock the step-by-step setup, a thumbnail prompt that works in any image AI, a filled-in example, and the prompt that turns it all into a Claude skill. Plus every other guide in the library. One email a week. Unsubscribe anytime.

Make your type sheet

One image that shows your font and your accent colour. Make it once in Canva, Figma or Keynote: a dark background, a sample title in your font, one word of it in your accent colour, and a swatch with the hex code. Save it as type-sheet.png. This is reference image 2 for every cover you make from now on.

Pick your title spot

Decide where the title lives and write it down in plain words, for example "top third, left-aligned, inside the top 40% of the frame". Keep it clear of the bottom of the cover, where Instagram's grid crop and the play icon can cover it. Use the same words in every prompt.

Grab a clean frame from the video

Find a moment where your face is sharp, your eyes are open and your mouth is closed. Pause there and screenshot it, or pull it at full quality with ffmpeg (change the timestamp to your moment). This is reference image 1.

ffmpeg -ss 00:00:12 -i video.mp4 -frames:v 1 frame.png

Run the prompt in any image AI that takes reference images

Upload the frame first and the type sheet second, then paste the prompt below with your placeholders filled in. It works in any model that accepts reference images, such as ChatGPT's image generation, Nano Banana or Higgsfield. In Midjourney, attach the two images as references. Each tool handles references a little differently, so if the face drifts, say "reference image 1" again in a follow-up and ask it to match the face exactly.

The cover lock prompt

Replace everything in [square brackets]. Keep the rest as it is.

Create a video thumbnail using the two reference images I've attached.

Reference image 1 is a frame from my video. Keep the person exactly as they appear: same face, same features, same skin tone, same hair, same clothes. Do not redraw, beautify or restyle them. You may crop, reframe and relight slightly so they read clearly at thumbnail size.

Reference image 2 is my type sheet. Match its typography exactly: the same font, weight, letter spacing and case.

Thumbnail spec:
- Aspect ratio: [ASPECT RATIO]
- Title text: "[TITLE TEXT]"
- Font: [FONT]
- Title colour: [BASE TEXT COLOUR]
- Highlight: only the word "[HIGHLIGHT WORD]" is in [ACCENT COLOUR]. Every other word stays [BASE TEXT COLOUR].
- Title position: [TITLE POSITION]. The title must sit in exactly this zone and must not overlap the person's face.

Style rules:
- Clean and high contrast. The title must be readable on a phone at grid size.
- Spell the title exactly as written. No extra words, subtitles, logos, watermarks, stickers, arrows or emojis.
- Use one accent colour only: [ACCENT COLOUR]. No other bright colours.
- Keep the background simple so the face and the title are the two things you see first.

Example: the prompt filled in

This is my setup: a clean bold sans, white title, one word in lime, top third, 4:5 for the Instagram grid.

Create a video thumbnail using the two reference images I've attached.

Reference image 1 is a frame from my video. Keep the person exactly as they appear: same face, same features, same skin tone, same hair, same clothes. Do not redraw, beautify or restyle them. You may crop, reframe and relight slightly so they read clearly at thumbnail size.

Reference image 2 is my type sheet. Match its typography exactly: the same font, weight, letter spacing and case.

Thumbnail spec:
- Aspect ratio: 4:5 portrait (1080×1350)
- Title text: "THUMBNAILS ON AUTOPILOT"
- Font: a clean, bold geometric sans-serif like Archivo Black, all caps
- Title colour: white (#FFFFFF)
- Highlight: only the word "AUTOPILOT" is in lime (#C6F432). Every other word stays white (#FFFFFF).
- Title position: top third of the frame, left-aligned, inside the top 40%. The title must sit in exactly this zone and must not overlap the person's face.

Style rules:
- Clean and high contrast. The title must be readable on a phone at grid size.
- Spell the title exactly as written. No extra words, subtitles, logos, watermarks, stickers, arrows or emojis.
- Use one accent colour only: lime (#C6F432). No other bright colours.
- Keep the background simple so the face and the title are the two things you see first.

Lock it in: save it as a Claude skill

Once a cover comes out right, save the whole setup as a skill so you never rebuild it. Open Claude Code in the folder where you keep your type sheet and paste this. Next time, ask for "a cover-lock thumbnail for my video about X" and it fills in the prompt for you. That's how cover 50 comes out looking exactly like cover 1.

Create a Claude skill called "cover-lock" in ~/.claude/skills/cover-lock/ that makes my video thumbnails consistent.

Store these as my locked settings in the skill:
- Font: [FONT]
- Base text colour: [BASE TEXT COLOUR]
- Accent colour: [ACCENT COLOUR]
- Title position: [TITLE POSITION]
- Aspect ratio: [ASPECT RATIO]
- Type sheet: the path to my type-sheet.png (copy it into the skill folder)

Also store this thumbnail prompt as a template with the same placeholders:
[PASTE THE COVER LOCK PROMPT HERE]

When I ask for a thumbnail, the skill should:
1. Ask me for the video file (or a frame) and the title. Suggest a title and a highlight word if I don't give one: 2 to 5 words, one word highlighted.
2. If I give a video, pull a clean frame with ffmpeg where my face is sharp, and show it to me before using it.
3. Fill the template with my locked settings, the title and the highlight word. Never change the locked settings unless I say so.
4. Give me the finished prompt plus the two reference images to upload: the frame first, the type sheet second.

Keep SKILL.md short and write it so it works the same way every time. Show me the finished skill when you're done.

Tips

Same order always upload the frame first and the type sheet second. The prompt calls them reference 1 and 2.
Short titles 2 to 5 words reads at grid size. Highlight the word that carries the promise.
Check the spelling image models still misspell text sometimes. Zoom in before you post, and regenerate if a letter is off.
Change nothing else the system works because the font, colour and spot never move. Change the title and the frame, nothing else.