Youtube Thumbnail

YouTube Thumbnail

Design thumbnails for high click-through on YouTube, for this niche: founders, agency operators, systems and AI tooling. The job is one second of attention on a 5-inch screen, in a feed, against nine competitors whose frames have already been measured.

Three decisions run the whole skill and none is a matter of taste:

1. The angle (step 1b). The claim the frame makes, the beat that proves it, and the title that claim wants. Decided before the picture, because it decides the video too.
2. Face variant or faceless variant (step 3). Decided by what the video is about.
3. How it is executed (step 7): designer handoff, real-pixel composite, or model generation. Decided by who executes and by whether a real person is in the frame. A face variant is composited from real pixels, never generated. The model draws a frame only when nothing real has to be preserved in it. See below.

Pick the template before you write a word of the frame (step 5c). Camera, ground, light, palette, depth and type placement live in a named template, never invented per run: that is why one set gave two good frames and four poor ones and the good ones taught the next set nothing. scripts/thumbnail-template.ts list, then references/templates.md. A concept naming none is unfinished.

The render-mode gate is the template's route, and it is mechanical now. On a face frame the only routes are photo, composite and edit-pass. Never a plain render-thumbnail.ts model run on a face frame: model mode hands the plate over as a reference and the model redraws him, so the frame reads as AI and is not him. This is the gate that gets skipped on a batch, which is why it is a field and a test now, not a paragraph.

Default deliverable: the build order in step 6, plus two A/B variants. Render only after the build order is approved.

The evidence this skill runs on

Everything below comes from 270 thumbnails, ten channels, each one split into a winner band and a control band from the same channel and period, so house style and production budget are held constant. A trait that appears in both bands is that channel's house style, and copying it buys nothing. Only traits that separate the bands are levers.

The findings, the counts and the per-channel table: references/niche-evidence.md. Read it before designing anything. It is the source of truth for this skill and it overrides any general "viral thumbnail" advice you already believe.

The three that overturn common advice, so they are repeated here:

• A face is not a lever. It appears in both bands on eight of nine channels. Presence buys nothing; a face-forward close-up filling the frame is a control marker.
• His face is never drawn. A generated frame contains zero photographed pixels of him, and it reads as AI at a glance. A face variant is always composited from the real plate: --photo for a full-frame plate, or a real-pixel composite when the background is a real artifact worth keeping (see references/composites.md). Handing the plate to the model as a reference is not compositing; it redraws him. See references/variants.md.
• Somebody else's money is a control marker. Client MRR, Stripe receipts, cash. Your own zero-to-X ladder is the one exception.
• Hype adjectives are a control marker. "INSANE", "UNLIMITED", "GAME CHANGER".

Part of this skill runs in code

Where an app writes the concepts rather than a person, its instructions are a runtime copy of the rules below, the compositions it picks from are a second file beside it, and the render path is a third copy of the execution rules. Change one and change the others in the same session.

Drift here is expensive and it is silent, because a generated concept looks finished. On 23 Aug 2026 the generator held the asset rules and the type rules and none of the composition evidence, so it wrote a frame carrying a face plus five vendor marks: six elements, which is the control-band shape this skill exists to avoid, arriving through the skill's own pipeline. It also named a shelved product. Both rules were in this file and neither was in the code.

When to use

• "Make a thumbnail for my video about X", "my CTR is low, redesign this thumbnail"
• "Give me 3 thumbnail variants to A/B", "write the brief so a designer can build it"
• Any request mentioning thumbnail, cover, CTR, video frame

Non-goals

• Channel banners, end screens, Shorts vertical covers (different aspect, different rules)
• Ad creatives and posters (use ad-copywriter)
• Long-form copy on the image

Workflow

1. Pull the brief

Get three things. Ask once, batched, and only for what context does not already give:

1. Video title, the exact words YouTube will show beside the frame
2. The promise in one sentence: what the viewer can do after watching
3. Who is making it: face available, brand, and the last three thumbnails published

The title is what the pipeline currently holds, not what has to ship. Step 1b decides whether it stays.

Do not ask for reference thumbnails. They exist on disk, already banded:

# build one contact sheet per channel: its 15 winners, its 12 controls

Ten channels are banded: matt-gray, ali-abdaal, chase-ai, liam-ottley, systems-made-better, nick-puru, ross-harkness, michele-torti, jordan-ross, and your own. Each writes sheet-winner.jpg and sheet-control.jpg. Read both. The read is already written up in references/niche-evidence.md; re-run the script only when the catalogues have been refreshed.

When the sheets leave the concept shapeless, or a set starts repeating the last set, read references/sourcing.md: the five reference sources beyond these nine channels, and how to take a reference's mechanism instead of its picture.

1b. Contest the angle

The claim, the beat that proves it, and the title it wants. Full doctrine, and why it is a step: references/angle.md.

• The claim is what a viewer believes the second the frame is read, in their words.
• The payoff is the beat that proves it, quoted from the script or the take. A claim nothing pays off is bait, and bait costs more than a weak frame: the click lands and the watch time does not.
• The title is held to the same law as the words on the frame. copy.md bans a category name and a sentiment on four words and used to wave them through on the sixty characters beside them.

This step runs before the script, which is the reason it is here and not at render time. A claim the video cannot pay off is either cut, or it is a beat the video gains while a beat still costs one bullet. The three concepts make three different claims about three different beats; three compositions of one claim is one concept drawn three times.

Where a take or a script exists, lift the claim out of it verbatim. A line he already says is a promise already paid.

The pick is recorded. Run with this angle on the thumbnails stage writes chosen onto the concept, and the script generator reads it as the debt the script has to settle in its first block.

2. Name the one thing in the frame

Write one sentence: "The eye lands on ___, and that tells the viewer ___."

If you cannot fill both blanks, there is no thumbnail yet. Winners in this niche carry one or two elements. Our own control band carries five to eight. That gap is the single biggest difference between our wall and theirs.

3. Pick the variant: face or faceless

Decided by the subject, not by preference. Full decision table and the composition rules for each: references/variants.md.

Both variants obey the same law: one person maximum, and never a second face. 1 winner in 135 has a second person; 12 of 108 controls do.

4. Write the words on the frame

Most of the CTR in this niche is won here, and it is the step that gets treated as a caption. Two to four words, naming the job rather than the feature, bounded by a count or a duration, and saying what the title does not. No adjective without an object.

The frame opens the loop on its own and the title is read second or not at all, so ask what the viewer should wonder and what they should feel before asking what the frame should show. Write five wordings, not one: the words are set in post from the concept's type block, so a rewording costs a screenshot rather than a generation.

The formulas, the winner and control verbatims they came from, the ban list and the full reasoning: references/copy.md.

5. Set the plate: contrast, type, safe area

Canvas 1280x720, judged at 320x180. One label block in one heavy condensed weight, two colours doing work, the bottom-right 15% clear of the timestamp and the bottom-left 10% clear of the progress bar. Two tests decide it: the 50% test (cover the bottom half, because that is all a TV row shows until the viewer presses down) and the squint test. The measurements and the Analytics check behind the first: references/craft.md.

5c. Fill the template

The frame is a template plus its slots, never a fresh composition. The five, and the rule that keeps the library honest (an empty shipped list means a proposal, not a template): references/templates.md.

5b to 8. Build, produce, execute, critique

The depth-and-light plate, the assets the frame is composited from, the designer handoff or the programmatic render, and the critique that runs before delivery are in references/production.md.

After it is published: the repackage loop

The title and the frame are the only two things a shipped video can still change, so this is the procedure for "my CTR is low". It runs on the video's rank in the channel's last ten by views, not on CTR: 1 to 3 leave it, 4 to 6 check the subject before blaming the frame, 7 to 10 repackage title and frame together and say the guess out loud first. The bands, the A/B margin that means anything, and why reviving a dead video means deleting it first: references/repackage.md.

The repair order across the whole video is subject, then title, then thumbnail, then hook. idea-mining owns the first, and it is the one that moves the number.

Reference

• references/angle.md - the claim, the payoff and the title, decided before the picture
• references/niche-evidence.md - what separates the bands, per channel and across channels
• references/sourcing.md - the five reference sources, and taking a mechanism rather than a look
• references/craft.md - depth, light and the floating object: why two frames holding the same things look different
• references/templates.md - the five templates, the slots they ask for, and how the library learns from a frame that shipped
• references/composites.md - producing the 3D tiles and artwork a frame is built from, and why prose never pins them
• references/anti-patterns.md - every measured control marker in one list, plus the three craft faults. Read it at step 8
• references/copy.md - the words on the frame
• references/repackage.md - the bands, the A/B margin, and reviving a dead video
• references/variants.md - face and faceless composition rules
• references/designer-handoff.md - the spec a designer builds from
• references/rendering.md - Nano Banana 2, models, keys, reference images
• references/archetypes.md - the seven general archetypes, and where this niche departs from them, plus references/viral-thumbnail-playbook.md for the general-YouTube background

Closing a run

A frame that shipped edits a template; a rejected one edits its negatives. Both go into thumbnail-templates.ts before the turn ends, with the date and the one thing the run taught. A fault recorded only in the pattern list is one the next generation makes again, because the generator reads the templates and not that file. Protocol: references/templates.md.

Learned Patterns

This skill appends new failure modes to its own pattern list after each run. They live in references/learned-patterns.md, newest first. Read that file before a run, and append to it after one whenever a run surfaces something not already there. A learning that stays in the conversation is lost when the conversation ends.

The run each line came from, with its quotes and numbers, is in references/learned-patterns-archive.md.