Analyse Thumbnails

---
name: analyse-thumbnails
description: Extract a channel's thumbnail style, your own or someone else's, before generating a new one. Triggered by "what are my thumbnails like", "analyse this channel's style", "what visual pattern does my competition follow". In Spanish, "cómo son mis miniaturas", "analiza el estilo de este canal", "qué patrón visual sigue mi competencia".
---

# Thumbnail style analysis

## WHEN

Proactive study of a channel's style, **with nobody having provided a specific
reference**. If the creator brings one thumbnail they like, that is
`thumbnail-inspiration`, not this skill.

It serves two purposes: knowing your own brand before generating something that
breaks it, and reading a competitor's visual pattern.

## EXECUTION ORDER

1. **Pick 3 to 5 representative videos**
   ```
   python3 tools/yt_recent_videos.py [--channel ID] --limit 15 --stats
   ```
   Representative of the **current** style: skip collaborations and
   off-brand experiments. If the channel changed style, only what came after.

2. **Download and measure the pixels**
   `python3 tools/yt_thumbnails.py --video ID1,ID2,ID3`
   Gives contrast, luminance, saturation and resolution. Objective, not
   arguable.

3. **Look at them.** Open each `file` path with the image reading tool. **This
   step is not optional**: the pixels do not tell you what is shown.

4. **Score one or two** to get the quantitative contrast:
   `python3 tools/score_thumbnail.py --video ID --title "..."`

5. **Extract the pattern**, noting in how many of the N each trait appears:
   - Colour palette and dominant temperature.
   - Typography: size relative to width, outline, word count. Flag it if the
     text occupies less than a third of the width.
   - Layout: where the subject falls, where the text does, what space is left.
   - Face usage: if one appears, **note the `video_id` where it reads best** —
     that is the likeness reference for generating later.
   - Recurring motifs: arrows, frames, split screens, logos, emoji.

   A trait appearing in 1 of 5 is not the channel's style, it is an exception.

6. **Cross against performance.** The pattern only matters against what worked:
   `python3 tools/yt_outliers_channels.py --min-ratio 2.0`.
   A consistent style that performs badly is a consistent style that performs
   badly.

## SOURCES

Pixel metrics `derived`. Views and outliers `derived` over `youtube_api`. The
style reading is the agent's judgement over images it has opened, and is
declared as such. The rubric is `heuristic`, unvalidated against CTR.

## OUTPUT

1. **A style brief in 5-8 bullets**, each with its frequency ("4 of 5").
2. **The 1-2 most representative `video_id`s** and why they are.
3. If there is a face, the `video_id` of the best likeness reference.
4. What must be **kept** so the brand is not broken, and what is **weak**
   measured against the rubric.