thumbnail
/thumbnail
Tworzy trzy bardzo różne miniatury 1280x720 do długiego wideo, z uzasadnieniem kliknięcia, testem nieznajomego i krótkim tekstem.
- Autor
- NatikChhabra ↗
- Repozytorium
- claude-arsenal ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 13.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 2
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Tworzy trzy bardzo różne miniatury 1280x720 do długiego wideo, z uzasadnieniem kliknięcia, testem nieznajomego i krótkim tekstem.
Wyzwalacze z opisu
Oryginalny opis (EN)
Make 3 drastically different high-CTR 1280x720 thumbnails for a long video (one Test & Compare pair each) - the click incentive named from the creator's own lines first, one audience per card, the stranger test, one felt focal point at phone size, a staged real frame, 0-3 words - then measured feed-size gates, fresh blind readers, a click floor and a harsh-scroller critique before his pick. Use for "thumbnail ideas", "make thumbnails", "my CTR is bad", or a Short/Reel cover.
Jak działa
- 0. Context first
- 0b. The click incentive, before any frame (video.md §1.1-1.7)
- 0c. The best-rated recipe: check every card (video.md §1.16)
- 1. Frames
- 2. The mascot (if the channel has one)
- 3. Compose 3 cards
- 4. Check (references/gates.md)
- 5. Deliver
- Covers for Shorts and Reels
Podgląd SKILL.md
---
name: thumbnail
description: Make 3 drastically different high-CTR 1280x720 thumbnails for a long video (one Test & Compare pair each) - the click incentive named from the creator's own lines first, one audience per card, the stranger test, one felt focal point at phone size, a staged real frame, 0-3 words - then measured feed-size gates, fresh blind readers, a click floor and a harsh-scroller critique before his pick. Use for "thumbnail ideas", "make thumbnails", "my CTR is bad", or a Short/Reel cover.
---
# thumbnail <ep>
status: beta
This skill is the CLICK method: the order of work, what to decide, which gate runs when, what gets logged. Every visual
rule and number (incentive table, hook lifts per audience, stranger test, focal point, words, measured floors, build
tiers, the lit cut-out recipe, brand mascot rules, never-on-a-card list, file spec, U1-U12, reader prompts) lives in
skill design, `references/video.md` §1, §5 and §6. Read the sections each step names. The gate run order and the
prompts not in video.md: `references/gates.md`.
## 0. Context first
Read the whole transcript (or the final caption file) and the video's concept note (premise, facts, beats). Research the
live context (collabs, news, trends around the topic and any creator the video references). Never ask the creator what
the recording already answers.
## 0b. The click incentive, before any frame (video.md §1.1-1.7)
A card set that was one template (a big moment + the mascot reacting + a caption you only get after watching) was
skipped by the creator 9 times in a row: no reason to click. Before any pixel, for each of the 3 pairs:
1. Read the cut's captions start to end; list the creator's lines that carry a number, a confession, a rule or an
absurd act. The incentive comes from his words; the monster or the scene is only the stage.
2. Pick ONE incentive per pair (§1.1), a different one each, and write the stranger's sentence: a verb, an open
outcome, TRUE in the cut (the line + its time). Store it with the pair (for example
`"incentives": {"<thumb>": "<sentence>"}` in the upload sheet).
3. List every card the creator rejected for this video: a new pair differs in the INCENTIVE, not only the frame or
the caption.
4. Name the video's 3 audiences by WHY they'd watch and give each card one (§1.2, hooks per audience §1.3). Never three
cards for one audience.
5. Run the stranger test on the picture alone (§1.4) and the focal-point check (§1.6, faces §1.7) on each idea.
6. Sketch the 3 ideas in words FIRST (subject, the question it leaves open, backdrop colour + why, words) and run the
T1 blind read (gates.md) on the sketch before composing.
## 0c. The best-rated recipe: check every card (video.md §1.16)
One line per rule per card; a "no" = rebuild: one face, one emotion a stranger names in one word, >= 60% of the height,
the sharpest thing on the card · that face does something TO the other subject (the creator's avatar is the loser;
beside is not a relation) · the stake in the topic's own look · <= 4 words from the cut's captions that can't be
misread · one backdrop colour, 3 elements, native-res crisp.
## 1. Frames
- First the frames shot on purpose at record time (tier 1, video.md §1.10), then the episode's biggest moments (the
edit's stings, loud spikes, the hook, any marker the creator said aloud). Grab ~8 candidates from the text-free
version of the cut: never a frame inside a motion graphic, meme or card span. The frame just BEFORE the sting is
usually the funny one; the picture leaves the question open (§1.5).
- Look at them; keep the 3 most readable at small size.
## 2. The mascot (if the channel has one)
Decide the pose yourself; never ask. Size, contact and the never-deform rules: video.md §1.12. Only the channel's own
approved mascot files. A missing emotion = a new pose made the channel's way, with the creator's approval.
## 3. Compose 3 cards
- 3 DRASTICALLY different ideas: a different hook mechanism, composition and palette each (§1.14). Layout free per
video; the channel's fixed identity (its text style, its mascot) stays.
- Lint each title + words pair: skill `yt-package`, `title.py --title "<t>" --thumb "<words>"`. Each thumbnail is built
for its own title (3 PAIRS per long, skill `publish-meta`).
- Finish and file: video.md §1.11, §1.13, §1.15 (a cut-out upscaled > 1.15x never ships; PNG over 2 MB = upload the
JPG twin).
- **Approved = frozen:** copy the approved cards to a `final/` folder and never re-render them through changed shared
code (an edge tweak re-rendered the creator's favourite and it came out worse). A tool change = render to a new name
and compare side by side.
## 4. Check (`references/gates.md`)
4a the measured floors first; 4b fresh readers: U1-U12, T1 + T4, T3, T5, T10 (a log), T12 click floor (>= 2 of 3) with
its incentive sentence, T11; then the harsh-scroller critique per card. Every reader is a FRESH agent given ONLY the
image, never the title (except T12 and the critique), the video or our reasoning.
## 5. Deliver
Show all three with the gate results (T1 sentence, T12 clicks, the critique lines) + one line each; his pick ships. Log
the pairs and, after day 7, the Test & Compare winner (watch-time share) with each variant's reader ranks: readers that
predict 3 winners in a row may become a gate.
## Covers for Shorts and Reels
The same method at 1080x1920; the bands, sizes and shelf rules: video.md §3.5. Its own T1 read at 202 px and a 9:16
feed mock with 3 readers; under 8.5 = redo. Details: skill `shorts`.
Pliki w paczce (2)
- SKILL.md 5 KB
- references/gates.md 4 KB
Wersje
W bazie nie ma innych skilli o nazwie thumbnail.
Migawka bazy z 13.11.2026 · wersja paczki 93e4127c955d (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100thumbnail-design
Projektuje miniatury YouTube, które zdobywają kliknięcia: czytelne na telefonie, z jednym punktem skupienia i uczciwe, wraz z briefem, tekstem na miniaturze i planem testów A/B według frameworka CLICK. - 100thumbnail-design
Projektuje miniatury YouTube czytelne na telefonie, z jednym punktem skupienia, obietnicą, tekstem i planem testu A/B. - 90creator-packaging
Projektuje miniaturę, tytuł i opakowanie wideo, które zwiększają CTR i wygrywają w feedzie, wraz z testami A/B. - 100thumbnail-design
Tworzy koncepcje miniatur YouTube, briefy projektowe i strategię poprawy CTR dla filmów, podcastów i kursów. - 100youtube-thumbnails
Projektuje i poprawia miniatury YouTube, przygotowuje warianty do testu A/B i ustala, który styl wygrywa na kanale, także gdy CTR wygląda dobrze, a test wybiera nudniejszy obraz. - 85youtube-packaging
Generuje pakiet YouTube nastawiony na kliknięcia: 3 propozycje miniatur pod jednym tytułem do wbudowanego testu A/B/C, opis nastawiony na wartość i renderuje miniatury jako obrazy.