_____ _____ .__
/ \ _____ ___ ____/ ____\_ __ _____|__| ____ ____
/ \ / \\__ \ \ \/ /\ __\ | \/ ___/ |/ _ \ / \
/ Y \/ __ \_> < | | | | /\___ \| ( <_> ) | \
\____|__ (____ /__/\_ \ |__| |____//____ >__|\____/|___| /
\/ \/ \/ \/ \/
A Claude skill that turns any article, script, or topic into ready-to-run thumbnail and cover-image prompts — with your own face kept consistent across every image.
Throw your content at the agent. It reads it, asks you three rounds of questions (format + style + title → reference images → visual details), then writes an image prompt engineered for your exact format — and either generates the images for you or hands you the prompts to run anywhere.
| Ratio | Built for |
|---|---|
| 16:9 | YouTube thumbnails, blog heroes |
| 9:16 | Reels / Shorts / TikTok covers, Stories |
| 1:1 | Instagram feed, podcast art |
| 5:2 | X (Twitter) Article cover headers |
| 3:4 | Portrait covers — Pinterest and vertical-feed platforms |
Every format has its own composition logic and safe zones (YouTube's duration badge corner, vertical platforms' UI bands, X's banner crop) — a prompt written for 16:9 is recomposed for 9:16, not just re-cropped. Pick several formats in one run and each gets its own adapted prompt.
All six generated with this skill — same face reference across every image, five formats, four styles.
 YouTube 16:9 — Dark Gradient |
 YouTube 16:9 — Poster Collage |
 Instagram 1:1 — Dark Gradient (narrative) |
 Portrait 3:4 — Dark Gradient |
 Reels 9:16 — Comparison Cards |
 X Article 5:2 — Dark Gradient |
At the end of every run you choose:
- Generate via MaxFusion MCP — the agent generates everything in chat. Default model Nano Banana Pro, automatic fallback to GPT Image 2, every MaxFusion image model selectable.
- Generate via your own API key — bring any image-generation provider's key (kept in your environment, never pasted in chat) and the agent runs the calls.
- Prompts only — you get the finished prompts and paste them into whatever tool you like. No keys, no dependencies.
| Style | Best for |
|---|---|
| Dark Gradient | Person centered, giant text behind, maximum impact |
| Flat Color | Clean cutout person + prop on a solid background |
| Product Hero | UI screenshot / product is the star; person guides the eye |
| Comparison Cards | Before/after, good/bad contrast content |
| Minimal Whitespace | Restrained, the title is the only focal point |
| Poster Collage | Many assets layered with depth |
| Side Subject | Person on one side, title owns the space |
| Back View | Person faces away — aspirational content |
| Partial Presence | Just a hand or half-face; product is the absolute star |
| Direct Gaze | Eye contact with the viewer, text framing the face |
Full prompt templates live in references/, with per-format adaptation notes in every style and 4 complete example prompts in references/examples.md as few-shot references.
- Image 1 is always your face photo — the skill keeps your facial features consistent across every generation (requires a model that accepts multiple reference images: GPT Image, Nano Banana Pro, Seedream 4.0…).
- Image 2+ is anything else going into the cover — product shots, UI screenshots, brand assets.
On first use the skill walks you through a one-time setup (default face photo + preferred delivery path), saved to a gitignored config.md so it never asks again.
Full clone required — the style templates and format spec live in references/; SKILL.md alone is incomplete.
git clone https://github.com/holy-templar/thumbnail-machine.git \
~/.claude/skills/thumbnail-machine
For the MaxFusion generation path, connect the MCP:
claude mcp add --transport http maxfusion https://mcp.maxfusion.ai/mcp
Then say something like: "make a YouTube thumbnail for this article" — and paste your content.
This skill is built on oh-my-cover-design by feitangyuan and its refinement gbro-cover-design by 狗哥笔记 (both MIT). What we kept, and what we changed:
Kept from the originals: the article-in → prompt-out concept; the 3-round batched question flow; on-demand style templates in references/; the few-shot example library; the face-as-image-1 convention; the first-run config wizard.
Changed:
- English throughout — image models render English text far more reliably than CJK glyphs, which removes most of the "check every character and regenerate" loop.
- 5 formats instead of one fixed 3:4 — each with its own composition logic and platform safe zones, plus multi-format batch runs.
- Style × format matrix — every style template carries per-format adaptation blocks (a Direct Gaze prompt at 5:2 puts text beside the face, not above/below it).
- Execution paths — the originals are prompt-only; this adds direct generation via the MaxFusion MCP (Nano Banana Pro default, GPT Image 2 fallback) and a bring-your-own-API-key path, while keeping prompt-only as a first-class option.
MIT — with upstream attribution to both parent projects.