YouTube Thumbnail Generator

Turns a thumbnail concept into rendered, verified frames — with the style built from an element menu you pick from, not a fixed template. One number matters: CTR.

Division of labor: /packaging decides what to bet on (one fixed title, 3 distinct thumbnail levers, honesty checks). This skill decides what the frame looks like and renders it. Run standalone for iteration ("give me a new B"), or let packaging Stage 5 delegate here. Both inherit two non-negotiables from packaging:

• Honesty guardrail — the frame must promise only what the video keeps.
• Loud-vs-calm — thumbnails are deliberately louder than brand.md. Never tone down to

match the in-video brand.

The baseline look (always on, never asked about)

Every render, regardless of selections: high contrast (deep darks OR clean brights, never flat) · vibrant saturated color with one accent doing the work · clear focal hierarchy (headline → hero object → face, three steps max) · big high-energy positive face from the media/library/faces/ kit · premium, clickable finish — crisp edges, no clutter · every text element perfectly spelled and razor sharp.

These came from a real before/after CTR comparison (the "Image 13" guide, 2026-08): the flat, low-contrast, no-hierarchy frame lost to the bold high-contrast hierarchical one ~4.4x on views. The axes are always-on; the decorations below are strictly opt-in.

The element menu (opt-in — this is what you ask about)

Exact prompt language for every element lives in references/style-elements.md. Summary:

| Element | What it is | Cost | |---|---|---| | single-hook | ONE text element, huge (word or number) | the default; cleanest | | tiered-headline | small white italic context line + huge green grunge power line | +1 text block | | sticker | tilted yellow badge, black caps, red underline, arrow to the hero | +1; ≤2 short words or it garbles | | checkstrip | slim bottom strip, 3 green-check benefits | +3 micro-texts; busy fast | | brush-tag | brush-stroke tag at strip end (e.g. WITH CLAUDE) | +1 | | results-fan | fan/burst of result cards from the hero (pure imagery, no lettering) | busy-frame risk | | glow-burst | neon energy + particles radiating from the hero | cheap, usually worth it | | real-logo | pass the actual logo file as a --ref so it renders faithfully | extra ref | | real-app-screen | pass a real UI screenshot as a --ref for the on-screen app | extra ref; micro-text caveat |

Environments (pick one): dark-studio (neon-on-black, dramatic) · real-office (candid DSLR look, anchored to the real room visible in the face refs) · bright-graphic (saturated studio composite, the classic loud style).

The flow

1 — Locate context

Project dir, existing packaging/thumbs/ (what letters/versions exist), the locked title if packaging ran (thumbnail text must never repeat it), face kit present (no kit → stop and ask, see media/library/faces/README.md), GEMINI_API_KEY in .env.

2 — The interview (the point of this skill)

Ask before generating — one AskUserQuestion round. Skip any axis the user already specified in their request; ask only what's still open.

1. Environment — dark-studio / real-office / bright-graphic.
2. Text budget — single-hook (recommended default) / headline-plus-one (tiered

headline + at most ONE accent element) / full-kit (headline + sticker + checkstrip + tag — warn that it's the busy end; opt-in only).

1. Extras (multi-select) — glow-burst, results-fan, real-logo, real-app-screen, a

prop/object, none.

1. Accent color — green (house default) / yellow / red / cyan.

Calibration note (real creator feedback, 2026-08): the default is MINIMAL text. The full-kit look was tried and rejected as "too much text"; high contrast + vibrant were explicitly kept. Never stack every element just because the guide contains them — each text block after the first must earn its slot.

3 — Confirm the words

Echo the exact hook text (and sticker/strip words if selected) before rendering. Check against the locked title: thumbnail text never repeats title words — the two combine into one message.

4 — Build the prompt

Assemble from references/style-elements.md blocks: labeled brief (Reference roles → Subject & Action → Composition & Hierarchy → Setting & Lighting → Text → Style), letter-by-letter spelling for every rendered word, positive framing with the one earned negation (sub-images carry no lettering/logos). Hardware rule: laptops/phones are "generic, plain dark bezel, no logos or lettering on the hardware" — otherwise you get a MacBook.

5 — Render

venv/Scripts/python tools/gen_thumbnail.py --prompt-file videos/<p>/packaging/thumbs/<X>.txt \
    --out videos/<p>/packaging/thumbs/<X>.png --jpg --seed <N>

• Default refs = the whole face kit. Any explicit --ref replaces that default — when

passing a logo or app screenshot, re-include the face ref explicitly, and name each image's role in the prompt ("Image 1 = identity, Image 2 = the app").

• real-app-screen source: pull a frame from the project's baked preview with ffmpeg, crop

to the window, save to media/projects/<p>/ (reusable, committed path) — never screenshot by hand if the beat already exists in the video.

6 — Verify (every render, before the user sees it)

Read the image back: hook spelled right + legible · face reads as the creator · one dominant focal element · bright/saturated/positive expression · hands sane (count fingers, no phantom limbs) · no stray lettering in sub-images · hardware unbranded · honesty (screen/props match the real product) · 16:9, JPG <2MB. Fix-and-rerender before surfacing; flag anything borderline honestly (e.g. micro-text garble in a real-app screen).

7 — Iterate

• Letters are bets, numbers are dressings (A, A2, D3…): a new concept gets a new

letter; a restyle/refinement of the same lever bumps the number. One A/B/C test slot per letter family.

• Before overwriting <X>.txt/.png, preserve the old one as <X>-v1.*.
• Hold --seed to iterate composition-stable; change seed to explore. Change one thing at a

time.

• Restitch the labeled contact sheet (thumbs/_ABC-set.jpg) after every accepted render so

the set is always comparable at browse-wall scale.

Failure modes → fixes (learned on real renders)

| Symptom | Fix | |---|---| | Sticker/badge word garbled | ≤2 short words per badge, letters spelled out (L-I-F-E-T-I-M-E); re-render | | Checkstrip drops a checkmark | phrase as "each item a green circular checkmark immediately followed by…" | | MacBook bezel / brand on hardware | "generic slim laptop, plain dark bezel, no logos or lettering on the hardware" | | Real app screen: micro-text garbles | expected — invisible at browse size; if it matters, blur-text reroll or perspective-warp the real screenshot in post | | Face color cast (purple/green hand or skin) | name the light: "warm skin tones preserved; the screen adds only a faint cool glow" | | Clone/double renders a stranger | "BOTH men in this image are this exact man" + give the double a distinct material (wireframe, hologram) | | Office looks generic | anchor it: describe the real room visible in the face ref (wall color, plant) and say "match the real office behind him in Image 1" | | Word spelled right but flat/pasted | grunge texture + soft outer glow + hard drop shadow, restate "razor sharp" |

Model ids, API details, template history: .claude/skills/packaging/references/thumbnail-generation.md. Worked examples (proven prompts): videos/video-1/packaging/thumbs/*.txt — A2 (bright-graphic + full kit + real logo), B (dark-studio outcome scene), C (dark-studio clone), D2 (real-office + full kit), D3 (real-office + single-hook + real-app-screen).