youtube-thumbnails

Overview

Turn a video's content into clickable packaging — title + thumbnail designed as one system, then rendered photorealistically from real headshots.

Core principle: packaging is psychology first, design second. A viewer's eye is caught by the thumbnail, checks the title for value, then returns to the thumbnail to confirm the promise — all in 1–2 seconds. Every step below serves that loop. A thumbnail that merely looks good but opens no curiosity gap will not get clicked.

Inputs to collect before starting

The content — transcript preferred (episode, video script, or detailed outline). No transcript yet? The yt-structure skill in this repo bundles scripts/yt-transcript to pull captions from a YouTube URL.
Headshots of everyone appearing — for the main subject, 2–4 well-lit photos with different expressions (shocked, skeptical, pointing…); guests need 1 clear front-facing photo (a LinkedIn photo works).
Style references — 1–2 of the channel's past thumbnails (the models replicate layout language, text-block style, and logo from these).
Gemini API key — in GEMINI_API_KEY or at ~/.gemini_api_key. Must be from a billing-enabled project (see Gotchas). Starting from zero: create a key at aistudio.google.com/apikey, click Set up billing on its project (image models have no free tier — a text-only key will 429), then save it: echo '<key>' > ~/.gemini_api_key && chmod 600 ~/.gemini_api_key.

Step 1 — Extract hooks and draft packaging concepts

Read the content and pull out 4–6 candidate hooks: surprising numbers, contrarian claims, named mistakes, before/after results. Then draft 4–5 title + thumbnail concepts, each built on a DIFFERENT hook — not five skins of one idea.