YouTube Thumbnail
20 original thumbnails for one long video, with your real face, in 20 proven styles, then the matched titles.
youtube-thumbnail makes 20 original thumbnails for one long-form video. Each
one is a different click idea, in a different style, with your real face. You
pick; nothing is ever uploaded to YouTube for you.
The styles come from a library of looks that work on YouTube, plus your own inspirations. Images are generated with Codex, free with your ChatGPT plan.
Invoke
Terminal
/youtube-thumbnail my video about Claude Code, the 30 day test
/youtube-thumbnail my-claude-code-test make a second, calmer batch
What it does
- First run only: sets up the profile. 1 to 5 sharp photos of you in
face/(front, three-quarter, smiling, surprised, no sunglasses, nobody else), and ayoutube.jsonwith your channel and how you look on camera (glasses, cap, beard) so the face stays consistent. A channel without a face can run in faceless mode. - Reads your video: transcript, chapters, notes, or your brief. A thumbnail never claims what the video does not say.
- Writes a strategy: what makes your thumbnails click (your own best ones first), then 20 click hypotheses, each with a viewer promise, a visual metaphor and 1 to 3 words of text in the video's language.
- Builds 20 concepts across the style library: channel classics, AI creators, clean editorial, creator inspirations and originals, at most 2 per style. A course video uses a dedicated mix.
- Generates the 20 images with Codex in the background (5 to 7 minutes), with the face, your identity line, the style reference and real logos injected into every prompt.
- Finalizes: 1280x720 JPG files in
miniatures/, contact sheets of 10 to browse, and a warning if two images are identical. The folder keeps every batch, with a prefix per batch. - Checks each one at full size: your real face (once), the exact text with its accents, real logos, no watermark, readable at phone size. A bad image is regenerated alone, never patched.
- Hands off: the 5 best are shortlisted, then youtube-title writes a title per thumbnail and the A/B trio.
What you need
- Codex installed and logged in with your ChatGPT account (see codex-images).
- Your face photos, once. Optionally your best thumbnails in
_inspirations/and real logos in_logos/. - You set the winner up yourself in YouTube Studio ("Test and compare").
Next
- youtube-title: the title for each thumbnail and the A/B trio.
- codex-images: the engine behind the images.
- publish: for shorts rather than long videos.