thumbnail-best-practices
Skill2 starsChanged 48 days agoMITView on GitHub
Thumbnail design rules — composition, text, mobile legibility. Triggered when generating or judging a thumbnail, whether your own or someone else's.
In this repository23 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- analyse-channel-packagingSkill
- analyse-thumbnailsSkill
- channel-analyticsSkill
- channel-identitySkill
- channel-kickoffSkill
- competitor-boardSkill
- competitor-ideationSkill
- competitor-mappingSkill
- competitor-scanSkill
- image-generation-coreSkill
- image-refinementSkill
- keywords-seoSkill
- likeness-preservationSkill
- memory-reviewSkill
- monetisationSkill
- packagingSkill
- script-writingSkill
- thumbnail-inspirationSkill
- transparencySkill
- youtube-banner-specSkill
- youtube-profile-specSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 50
Only what actually appears in the video.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Thumbnail design rules
WHEN
Generating a thumbnail, diagnosing one, or extracting a channel's style. These are design rules: they apply regardless of which provider generates.
RELATIONSHIP WITH THE CHANNEL'S RUBRIC
These rules are general. config/rubrics/thumbnails.yaml is the rubric
calibrated with this channel's memories, and it wins where they conflict.
Its judgement axes carry 60 of the 100 points:
not_redundant_with_title(25) — if the hard number is in the title, the thumbnail shows the visual result, it does not repeat the number.result_not_ui(18) — a timeline or a dense interface is grey noise at feed size.logos(9) — one at most, and only if it is an anchor your audience recognises.short_text(8) — four words or fewer.
On a fresh install those four axes are uncalibrated examples: each one's
memory field is null. Adapt them to your channel and, when you learn
something from your own thumbnails, write it into memory/ and point the axis
at that file. When generating, the live axes go into the prompt.
GENERAL RULES
Mobile legibility — the real bar. The thumbnail is seen at roughly 160 px wide.
- The text should occupy a third of the width or more. Below that it is illegible.
- Heavy bold sans-serif, thick outline or hard shadow.
- A simple background area behind the text, for maximum contrast.
- Three to five words maximum.
Composition
- Rule of thirds: the focal point on an intersection.
- Negative space reserved for the text, preferably at the top.
- Leading lines (arms, arrows, architecture) pointing at the subject.
- Bottom right corner left clear: the duration stamp goes there.
- One single focal idea. Limited palette. High subject/background contrast.
Authenticity
- Only what actually appears in the video. A scene that does not exist is clickbait, and it penalises growth even when it lifts clicks.
- Realistic aesthetics over over-processed plastic.
- The expression should match the video's real emotion.
- Show the action, not a static pose.
Human elements
- A face helps but is not mandatory: hands, POV or an object work too.
- Different clothes and pose in each video, to signal fresh content.
What to avoid
- Crowded composition, small text, stock-photo aesthetics, thick borders, mixed colour temperatures (warm orange with daylight).
VERIFICATION
Measure what is objective, do not opine about it:
python3 tools/yt_thumbnails.py --video ID # contrast, saturation, resolution
python3 tools/score_thumbnail.py --video ID --title "..."
python3 tools/score_thumbnail.py --image path.jpg # for one just generatedscore_thumbnail scores the automatic 40% and returns the 60% as unanswered
questions. You answer them by opening the file. Inventing that 60% is the
failure the tool is designed to prevent.
SOURCES
The rubric is heuristic and not validated against CTR while
data/history/studio_ctr.json does not exist. A score is never presented as
a click prediction. yt_thumbnails' pixel metrics are derived.
OUTPUT
- The measured part first:
score_thumbnail's automatic 40% andyt_thumbnails' pixel metrics, labelled with theirsource. - The judgement axes answered after opening the image, one line each.
- The two or three rules above that the thumbnail breaks, each with the concrete change that would fix it.
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.

