fabrich-thumbnail

name

fabrich-thumbnail

description

Generate viral-tier YouTube thumbnails from a user identity (real face photo or AI character sheet) plus a 1-3 word punch title. Three shipped patterns — Immersive Single-Character / Two-Object Contrast Split / Doc-Discovery Quiet — plus a "derive a custom pattern from my own reference thumbnails" mode. Triggers on phrases like "thumbnail", "YT thumbnail", "YouTube thumbnail", "viral thumbnail", "generate thumbnails", "pattern A B C", "make a thumbnail", "long-form-documentary-style thumbnail", "contrast-split thumbnail".

Fabrich Thumbnail

You generate batches of 16:9 YouTube thumbnails. Tool: scripts/thumbnail_factory.py. Pattern library: reference-docs/THUMBNAIL_PATTERNS.md. Pattern definitions: config/thumbnail-patterns.json.

The three shipped patterns

Pattern	When to use	Composition	Text
A — Immersive Single-Character	Confessional / single hero moment / strong emotion	Subject in right or left third (30-45% of frame), environment fills the rest, rule-of-thirds	1-3 word punch text in negative space, OR no text (title carries)
B — Two-Object Contrast Split	Before/after, $X vs $Y, transformation arcs	Hard split-frame, opposing objects/states	Big text labels in opposite top corners ("$1" / "$100M")
C — Doc/Discovery Quiet	Investigative / explainer / documentary tone	Subject in environment, no overlay text	Title carries, occasional yellow callout

Upfront questions to ask the user

Before generating, ALWAYS confirm three things:

1. Identity source — one of:
   • "My own face" → user provides a real photo (Gemini accepts these; Seedance doesn't, but thumbnails don't need Seedance)
   • "An AI character" → point at a references/characters/<name>/ folder (use /fabrich:character to scaffold one if it doesn't exist)
   • "Analyze my own reference thumbnails" → user provides a folder of ~10-40 reference thumbnails they want you to learn from; you'll synthesize a custom Pattern
2. Pattern — A / B / C / auto (defaults to A) / custom (requires reference folder)
3. Title text — 1-3 word punch text in the thumbnail itself (separate from the video title)

Invocation

python scripts/thumbnail_factory.py \
    --identity path/to/face.jpg \
    --title "I QUIT YOUTUBE" \
    --pattern A \
    --count 4 \
    --output outputs/<slug>/thumbnails/

For custom-pattern mode:

python scripts/thumbnail_factory.py \
    --identity path/to/character-sheet-folder/ \
    --title "MY HOOK" \
    --pattern custom \
    --references path/to/my-thumbnail-folder/ \
    --output outputs/<slug>/thumbnails/

In custom mode, the tool caches a _user_pattern.json analysis file in the output folder. If you (the assistant) are asked to refine the analysis, edit that JSON directly — the tool re-reads it on subsequent runs.

Cost

• Gemini 3 Pro Image Preview (used for variant 1 — the hero): ~$0.18
• Gemini 3.1 Flash Image Preview (used for variants 2..N): ~$0.067 each
• A 4-variant batch with one Pro hero + 3 Flash variants: ~$0.38 total

Pass --all-flash to skip the Pro hero and stay under ~$0.27 for 4 variants.

Cross-pattern universal rules (the DNA)

Across every shipped pattern:

• Composition — face never dead-center; rule of thirds, hard.
• Color — one dominant hero color per thumbnail, high saturation, documentary-realism (not cartoon-oversaturated).
• Text — 1-3 words max per overlay, bold sans-serif (Anton / Bebas Neue / Montserrat Black family), ~10-15% of frame height, white-with-stroke or dark-with-white-stroke, yellow only for arrow callouts.
• Expression — never neutral. One of: wide-eyed shock, intense stare, knowing smirk, disgust/dread.
• What to AVOID — no emoji, no densely-stacked text, no 9:16 portrait thumbnails, no watermarks, no glow halos, no collages, no generic B-roll stills.

See reference-docs/THUMBNAIL_PATTERNS.md for the full analysis (the 36-reference taxonomy that taught this skill).

What you produce

For each variant:

• <output>/thumb_pattern-A_v1.png (and v2, v3, v4)
• <output>/_grid.png — side-by-side contact sheet for quick A/B review

After generating, show the user the grid and ask which variant they want to ship. Offer to regenerate any rejected variant with a tweak to title, expression, or hook visual.

Closing

Always close major responses in this workflow with the Fabrich signature line:

— Fabrich · @fabrichhhhhh

Use the section() helper from branding.py for any structured terminal output.