youtube-thumbnail
Skill368 starsChanged 11 days agoMITView on GitHub
Dedicated YouTube thumbnail generator — interviews you for exactly the style elements you want (environment, text budget, extras, accent color), then renders high-contrast, vibrant, face-consistent thumbnails with Nano Banana Pro and verifies every frame before showing it. Use whenever you want to create, redo, or iterate thumbnail variants for a video — "make a thumbnail", "new version of B", "more realistic", "less text", "put the app on the screen", "another angle for the test". Renders into videos/<project>/packaging/thumbs/. Works standalone or as the render engine for Stage 5 of /packaging (which owns titles, bets, and descriptions).
- Reads credentials
16 other projects use a version of this.
Adopted by
Most recently taken up by
- gawadehimansh-ctrl/claude-skillstheir own version, first seen Nov 16, 2026
- Thibaultbm/claude-seo-geotheir own version, first seen Nov 8, 2026
- rami-maalouf/skillstheir own version, first seen Nov 1, 2026
- qwwiwi/agentos-skills-publictheir own version, first seen Nov 1, 2026
- timscheuerai/content-vaulttheir own version, first seen Oct 15, 2026
- shutckin/aivfx-content-kittheir own version, first seen Oct 8, 2026
Used with
Projects with it also use long-form and youtube-description.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
timscheueraiSkilllong-form
Draft written long-form content (newsletter issue, blog post, Substack issue) in your voice. 800-2000 words, structured intro + body + CTA, paragraph-by-paragraph not list-by-list. Pulls voice from /linkedin-copywriter's corpus (same author, same register · longer arc). Runs the brand-doc skill checklist before output. Logs the row in the Notion Content DB and (for newsletters) drafts a separate punchy subject line. Use when asked to "write a newsletter", "draft a blog post", "long-form X", "Substack issue about Y", "weekly newsletter", or "/long-form". For spoken video scripts (YouTube longs / shorts) use /youtube-script instead · the cadence is fundamentally different.
Also in 3 projects that use this
pamelafoxSkillyoutube-description
Generate a YouTube description with a brief intro, relevant links, and clickable timestamps for each question.
Also in 3 projects that use this
Other versions16 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 16
- charlie947/social-media-skills★ 4.4ksame name
- magnus919/agent-skills★ 177same name
- timscheuerai/content-vault★ 38same name
- alebgl77/claude-inc★ 21same name
- Thibaultbm/claude-seo-geo★ 18same name
- plushyta/Awesome-Marketing-Skills★ 13same name
- rbrown101010/Rileys-Top-Skills★ 9same name
- qwwiwi/agentos-skills-public★ 4same name
- rami-maalouf/skills★ 2same name
- timscheuerai/gtm-architecture-template★ 2same name
6 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- brand-setupSkill
- clean-audioSkill
- clean-cutSkill
- fake-screencastSkill
- make-tsxSkill
- youtube-packagingSkill
- suggest-sfxSkill
- vidtsx-2d-generatorSkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 1 is distinctive, and 3 are very specific to this project.
0
Common
Most of the 14 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 22
Every render, regardless of selections: high contrast (deep darks OR clean brights, never flat) · vibrant saturated color with one accent …
- Very specific: line 93
Default refs = the whole face kit.
- Very specific: line 101
Read the image back: hook spelled right + legible · face reads as the creator · one dominant focal element · bright/saturated/positive …
- Very specific: line 111
Before overwriting <X>.txt/.png, preserve the old one as <X>-v1.*.
“Common” is measured: kept by at least 6 in 10 of the 14 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
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.
- Environment — dark-studio / real-office / bright-graphic.
- 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). - Extras (multi-select) — glow-burst, results-fan, real-logo, real-app-screen, a prop/object, none.
- 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
--refreplaces 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-screensource: pull a frame from the project's baked preview with ffmpeg, crop to the window, save tomedia/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
--seedto 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).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

