thumbnail-machine

---
name: thumbnail-machine
description: >-
  Generate ready-to-run thumbnail/cover image prompts from any article, script, or topic — 5 formats
  (16:9, 9:16, 1:1, 5:2, 3:4), 10 composition styles, real-person face consistency via
  multi-reference images. Three delivery paths chosen at the end: generate directly via MaxFusion AI
  MCP (default model Nano Banana Pro, fallback GPT Image 2), generate via the user's own image-API
  key, or prompts only. Use whenever the user wants a thumbnail, cover image, YouTube thumbnail,
  X article cover/header, Instagram cover, Reels/Shorts/TikTok cover, blog hero image, or says
  "make a thumbnail", "cover for this article", "thumbnail machine".
---

# Thumbnail Machine

Flow: read the user's content → three rounds of questions → read the format spec + the chosen style template → produce the prompt(s) → delivery gate (MaxFusion MCP / own API key / prompts only).

---

## First run: config check

Every time this skill triggers, check whether `config.md` exists in this skill's directory:

- **Exists** → read it, go straight to the normal flow.
- **Missing** → first use. Run the one-time setup below, then continue.

### One-time setup (ask everything at once)

Explain how the skill works first: it turns content into cover-image prompts; generation is optional and happens only at the final delivery gate. Then ask three things:

1. **Default face reference** — covers feature a real person, which needs one clear front-facing photo used as image 1. Two options:
   - Save one to this skill's directory as `assets/my-face.png` (recommended: auto-used as image 1 on every run)
   - Choose "upload each time"
2. **Preferred delivery path** — MaxFusion AI MCP (generate in chat), own API key, or prompts only. This becomes the default; the gate still confirms each run.
3. **If prompts-only or own key:** confirm their image model supports **multiple reference images** (image 1 face + image 2+ assets) — e.g. GPT Image, Nano Banana Pro, Seedream 4.0. Single-reference models cannot hold face consistency.

Write the answers to `config.md` in this skill's directory (it is gitignored and never committed):

    # Config
    - Default face: assets/my-face.png (or: upload each time)
    - Delivery: maxfusion-mcp | own-key (<provider>) | prompts-only
    - Configured: <YYYY-MM-DD>

---

## Question flow (three rounds — ask each round's questions together, never one at a time)

### Round 1 — Format + style + title

Read the user's content first. Then, in one message:

**Format.** If the user already named a platform or ratio ("YouTube thumbnail", "X article cover"), skip the question and state the inferred format. Otherwise present:

| # | Ratio | Typical use |
|---|-------|-------------|
| 1 | 16:9 | YouTube thumbnail, blog hero, web |
| 2 | 9:16 | Reels / Shorts / TikTok cover, Stories |
| 3 | 1:1 | Instagram feed, podcast art, square social |
| 4 | 5:2 | X (Twitter) Article cover header |
| 5 | 3:4 | Portrait covers — Pinterest, RedNote/WeChat-style vertical |

Multiple formats may be selected — each gets its own format-adapted prompt in the same style.

**Style.** Recommend 1–2 styles for this content with a one-line reason each, then show the full menu:

| # | Style | Best for |
|---|-------|----------|
| 1 | Dark Gradient | Person centered, giant text behind, maximum impact |
| 2 | Flat Color | Clean cutout person + prop on a solid background |
| 3 | Product Hero | UI screenshot / product image is the star; person guides the eye |
| 4 | Comparison Cards | Before/after or good/bad contrast content |
| 5 | Minimal Whitespace | Restrained, text is the sole focal point |
| 6 | Poster Collage | Many assets layered with depth |
| 7 | Side Subject | Person on one side, title owns the open space |
| 8 | Back View | Person faces away — aspirational, new-beginnings content |
| 9 | Partial Presence | Only a hand / half-face shows; product is the absolute star |
| 10 | Direct Gaze | Eye contact with the viewer, text framing the face |

**Title.** Propose 1–3 title candidates extracted from the content. Length by format: 16:9 and 9:16 → 3–6 words; 5:2, 1:1, 3:4 → up to ~8 words. The user picks or types their own.

### Round 2 — Reference images

Ask both together (skip question 1 if `config.md` sets a default face and `assets/my-face.png` exists — just state it will be used as image 1):

1. **Image 1 (face reference):** have one → upload it; the prompt will say "match the facial features of image 1". None → ask for a short description of the person (gender, general look) and describe them directly in the prompt.
2. **Extra assets (image 2, 3, …):** UI screenshots, product shots, logos, other reference images. Have some → how many, and what is each? None → skip.

### Round 3 — Visual details

List all four items at once; the user replies by number (e.g. "2 / 1 / 6 / 3"). Anything not answered defaults to "let the model decide". Per-style skips: Back View skips A; Minimal Whitespace and Side Subject default to light backgrounds, so ask B only as "which light tone".

**A. Expression**

1. **Hands-over-mouth shock** — both hands over mouth, wide eyes
2. **Open-mouth shock** — jaw dropped, eyes huge
3. **Big laugh** — genuine joy, eyes crinkled
4. **Excited lean-in** — leaning forward, barely contained energy
5. **Confident smirk** — slight smile, steady eyes, "it's that easy"
6. **Chin-rest thinking** — hand on chin, contemplative
7. **Recommending nod** — warm smile at camera, "you have to try this"
8. **Let the model decide**

**B. Background tone**

1. **Light** — white / off-white / light gray, clean
2. **Dark** — charcoal / ink black, weighty
3. **Warm** — cream / warm brown / soft orange
4. **Cool** — blue-gray / teal-gray, calm and professional
5. **High-saturation clash** — model picks one high-contrast color pair, energetic
6. **Let the model decide**

**C. Font style**

1. **Ultra-bold sans** — heavy and forceful; tech / how-to first choice
2. **Soft rounded** — rounded strokes, warm and friendly
3. **Handwritten scribble** — natural strokes, personal
4. **Minimal sans** — even strokes, editorial, designed
5. **Classic serif** — serifed, literary / nostalgic
6. **Clickbait block-highlight** — viral short-form title style; full spec at the end of this file
7. **Let the model decide**

**D. Text color treatment**

1. **Pure white** — first choice on dark backgrounds
2. **Pure black** — first choice on light backgrounds
3. **Gradient** — model matches a gradient to the background tone
4. **Outlined** — stroked outline for extra separation
5. **Let the model decide**

---

## Writing the prompt

After the three rounds:

1. **Always read `references/formats.md`** — it defines the mandatory first line, composition logic, and safe zones for the chosen format(s).
2. **Read the chosen style template**: `references/style-XX-<name>.md` (XX = 01–10), including its **Format adaptations** section.
3. **Read `references/examples.md`** and use the closest example as a few-shot reference for *specificity* — poses written down to where each hand is, props described down to what is printed on them, spatial relationships stating what occludes what.
4. Fill the template.

### Core requirements

- **Poses are specific**: body position, where each hand is, what the hands are doing — never "she gestures".
- **Main visual elements are detailed**: what it looks like, what's written/shown on it, relative size, static or in motion.
- **Spatial relationships are explicit**: foreground / midground / background, and which element occludes which.
- **Backgrounds and fonts differ per style** — never reuse one color/font recipe across styles.

### Universal rules

1. **First line = the format line** from `references/formats.md`, verbatim (e.g. "16:9 widescreen composition."). One prompt per selected format.
2. **Safe zones** per format as defined in `references/formats.md` — they differ; do not default everything to 10% margins.
3. **Title text goes in quotes** in the prompt. After generation, proof every letter — image models still misspell. On a typo: **report the defect and ask the user whether to regenerate. Never regenerate on your own initiative** — every rerun costs the user money.
4. **Concrete metaphors** — abstract concepts become recognizable physical objects.
5. **No glow effects** — no holograms, data streams, or particle effects; they read as cheap.
6. **Positive description only** — write what should appear, not what shouldn't.
7. **Reference images are called "image 1 / image 2 / image 3"** — never re-describe their full contents.
8. **One style per image** — never blend styles.

---

## Delivery gate

When the prompt(s) are written and shown, ask exactly:

> **How do you want this produced?**
> 1. **Generate now via MaxFusion AI MCP** (default)
> 2. **Generate via your own image-API key**
> 3. **Prompts only** — copy them and run them anywhere

If `config.md` records a default path, propose it first but still confirm.

### Path 1 — MaxFusion AI MCP

1. Run a **live capability check**: list the current image models on the MaxFusion MCP with real calls. Never assume the catalog from memory.
2. Model choice: default **Nano Banana Pro** (`nanobanana-pro`). If it is not in the live catalog — or doesn't offer the needed aspect ratio — fall back to **GPT Image 2**. If the user names another available model, use that. All MaxFusion image models are selectable.
3. Upload image 1 (face) and any extra assets as reference inputs.
4. One generation per selected format, at that format's aspect ratio.
5. Poll each job to completion and deliver every image **in chat as inline widgets** — never raw storage links, no canvas.
6. Proof the rendered title text. On any misspelling or defect: show it, name the defect, and ask before regenerating — never rerun unprompted.

### Path 2 — User's own API key

1. Ask which provider (OpenAI, Google, FAL, Replicate, …) and confirm the key is available to the agent's environment (e.g. an env var or local `.env`) — **never ask the user to paste a raw key into chat**.
2. Confirm the chosen model supports **multiple reference images**; if not, warn that face consistency cannot be guaranteed and offer to proceed without image 1.
3. Call the provider's image API directly with the prompt(s) and reference images, one call per format, and return the results in chat.
4. Same text-proofing reminder.

### Path 3 — Prompts only

Deliver, per selected format, one copy-paste block containing the final prompt, followed by run instructions:

- Attach the face photo as **image 1** and any assets as image 2+ (multi-reference model required for face consistency).
- Set the aspect ratio to the format's ratio.
- Paste the prompt, generate, and proof the title text letter by letter.

---

## Hard rules

- Never generate before the delivery-gate answer.
- The format line is always the first line of every prompt.
- Never ask questions one at a time — rounds are batched.
- Never put an API key, or any secret, into any file in this skill.
- Multi-format runs: same style and content across formats, each prompt adapted per `references/formats.md` — never one prompt reused verbatim across formats.

---

## Font style appendix: Clickbait block-highlight

When the user picks font style 6, the font description in the prompt must read:

Viral short-form title style: ultra-bold condensed sans-serif, squared and heavy (in the spirit of Archivo Black / Anton), enormous type filling its zone, tight line spacing with slightly widened letter spacing for a pressurized, high-CTR look. The 1–2 most important keywords sit on high-saturation bright-yellow rectangular highlight blocks with black ultra-bold text — like warning signs, impossible to ignore.