/thumbnail-maker

name: thumbnail-maker
description: "Produce YouTube / blog / podcast-episode thumbnails. 16:9 default (1280×720, 1920×1080). Face + bold title aesthetic; face-placement variants (left / right / center). Model picked for face preserve + text rendering. Use when: 'YouTube thumbnail', 'blog header image', 'podcast episode cover', 'обложка для youtube видео / поста / эпизода'."

license: MIT
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Grep
  - Glob

<objective> Thumbnail generator. Input: title + optional face photo + optional style. Output: N variants at 16:9 (YouTube / blog / podcast-episode thumbnail).

Distinct from `cover-maker`:

• Aspect: ALWAYS 16:9 (no medium variation)
• Aesthetic: bright contrast, large text, exaggerated facial expressions — "thumbnail look"
• Variants vary in FACE PLACEMENT (left / right / center) more than in style

Distinct from `flyer-maker`:

• No event details (date / location)
• Stronger emphasis on the FACE if photo provided
• Different typography conventions (bigger, often colored, with stroke / shadow)

This skill does NOT:

• Generate YouTube banner art (different aspect — use `cover-maker --aspect 2560x1440` or `image-prompt --execute`)
• Generate animated thumbnails — static images only
• Source rights-cleared face references — provide your own
• Optimize for click-through (subjective — picks best of 3-5 manually)

</objective>

ROLE

Read title + optional face → pick model (text-friendly + ref-capable) → assemble per-variant prompts with face-placement + bold title composition → batch execute → save PNGs.

PIPELINE

1. **Resolve title**:

• Required: `--title`
• 3-7 words is the sweet spot for thumbnails
• High-energy phrasing works best ("How I Built X in 30 Days", "STOP DOING THIS")

2. **Resolve photo** (optional):

• `--photo <path-or-url>` — face reference
• Identity preserve via Nano Banana Pro (default)

3. **Resolve style** — see `references/composition-zones.md`:

• `--style auto`: picks vibrant + text-friendly styles
• Default for "YouTube look": `gradient-mesh-modern` or `swiss-grid-poster` with vibrant accent

4. **Face placement variants**:

• `--placements left,right,center` — default `left,right,center` (3 variants)
• The same face / title / style; placement varies
• Maximizes the chance of "the one that works for click"

5. **Pick model**:

• Photo + face → `nano-banana-pro` (identity preserve)
• No photo + heavy text → `ideogram-3-quality`
• Mixed → `gpt-image-2`

6. **Build per-variant prompts** — composition zones per placement (face-left + text-right; face-right + text-left; face-center-bottom + text-top).

7. **Batch execute** — parallel.

8. **Output**:

   ./generated/thumbnail/<slug>/
     thumbnail-left.png       (face left, text right)
     thumbnail-right.png      (face right, text left)
     thumbnail-center.png     (face center, text top)
     manifest.json
     style-used.md
     prompts.md

MODES

Required

• `thumbnail-maker --title "<text>"`

Optional

• `--photo <path-or-url>` — face / subject reference
• `--subtitle "<text>"` — secondary line (rare for thumbnails)
• `--style auto|<library-id>` — visual style
• `--style-mod "<override>"` — tweak
• `--placements left,right,center,top,bottom` — comma list (default `left,right,center`)
• `--variants N` — variants per placement (default 1; >1 gives N takes per placement)
• `--lang en|ru` — language hint
• `--model auto|<slug>` — provider
• `--aspect WxH` — override default 1920×1080 (e.g., `1280x720` for low-res)

Type presets

• `--type youtube` — 1920×1080 (default), bright + bold aesthetic
• `--type blog` — 1200×630 (OG image standard), more editorial
• `--type podcast-episode` — 1920×1080 with podcast-cover-feel

Execution

• `--execute` — actually generate
• `--output <dir>` — custom output
• `--parallelism N` — concurrent calls
• `--yes` — skip cost confirmation
• `--resume` — retry failed
• `--prompts-only` — dry run

REFERENCES (load on demand)

| File | When to load | |---|---| | [references/model-picker.md](references/model-picker.md) | Step 5 — model auto-pick + capability matrix | | [references/composition-zones.md](references/composition-zones.md) | Step 6 — per-placement composition templates, face + text positioning | | [references/troubleshoot.md](references/troubleshoot.md) | When face placement collides with text, contrast issues, click-bait avoidance |

EXAMPLES

See [examples/before-after.md](examples/before-after.md) — 3 calibration runs: YouTube tutorial thumbnail with creator face, blog header without photo, podcast episode cover with guest portrait.

CONSTRAINTS

• **Title 3-7 words.** Past 7, text crowds + becomes unreadable at thumbnail scale.

• **Title is THE focal point.** Even with a face, the title must be readable at 320×180px (small thumb on mobile).

• **One face per thumbnail.** Multiple faces compete; pick one main subject.

• **Face placement variants are stochastic.** "Left" / "right" / "center" guide the model but don't lock pixel-perfect.

• **Bright contrast is the YouTube look.** The default `--style auto` biases toward vibrant + high-contrast styles. Override with `--style photo-editorial-bw` for editorial blog feel.

• **Don't add too many text elements.** Title + optional subtitle = max 2. Past that, thumbnail looks crowded.

• **Reasonable click-bait avoidance.** The skill doesn't enforce "non-clickbait" — that's editorial judgment. But it also doesn't artificially exaggerate facial expressions unless you ask via `--style-mod "exaggerated shock expression"`.

• **One model across the variant set.** Lock per-thumbnail.

• **Cost confirm ONCE.** Sum across placements × variants.

• **Per-platform image sizing**:
• YouTube: 1280×720 minimum, 1920×1080 high-res
• Blog OG image: 1200×630
• Podcast episode art: 1920×1080
• LinkedIn article cover: 1200×627

• **Never print API keys.** Mask in errors.

INVOCATION HINTS

When th