Avenox Thumbnail

Thumbnail Factory

A systematic pipeline for producing scroll-stopping YouTube thumbnails for a mascot-driven channel. It encodes the formula, the working prompt recipe, the hook-pattern playbook, and the hard safety rules learned in production.

Built for anonymous / faceless channels: the mascot is the face, so the covers stay consistent without ever putting a person on screen.

Anonymity rule: never put the creator's real name, personal handle, or likeness on a thumbnail. If the channel is anonymous, the mascot is the identity — breaking that once is not reversible.

Configure before first use

The formula

Every thumbnail = 4 ingredients, nothing more (3-element rule; blink-test first — if it doesn't read in a quarter second at sidebar size, it failed):

1. The mascot — on-model, expressive pose that fits the video's emotion.
2. One 2-4 word hook — MASSIVE bold sans-serif caps, white + ONE accent word, thick black outline. Never duplicate the video title; open a curiosity gap that the title and video close.
3. Relevant tool logo(s) — so the topic is legible in one glance.
4. High-contrast modern background — dark + one strong accent color (red = urgency, green = success, blue = tech, purple = automation, gold = money). Bright, saturated, clean.

Design levers (apply, don't just decorate): curiosity gap • negativity/risk • contrast (before/after, ✓/✗) • tangible number or timeframe • authority • one clear focal point • lower-right corner kept clear (YouTube stamps the duration there) • text left / mascot right reads best.

Full hook playbook → references/hook-patterns.md.

Standard workflow

1. Read the video — title + description/topic. Identify the single most clickable angle.
2. Propose hooks (unless told "just fire") — 2-3 distinct concept directions: hook words + composition + which logo + which pattern. Let the user pick, or fire all as alternatives.
3. Fire in PARALLEL — one codex exec gpt-image-2 job per concept, all run_in_background: true in a single message. The user wants alternatives; compute isn't the constraint. Recipe → references/prompt-template.md.
4. Save all outputs to $THUMB_OUT with descriptive names (<video>-<letter>-<hook>.png).
5. md5-dedupe after the batch — parallel gpt-image-2 jobs can occasionally return duplicate images. Re-fire any duplicate solo.
6. Show + recommend — look at each PNG, give an honest verdict, say which title it pairs with, and make an A/B recommendation.
7. Deliver finals — sips -z 720 1280 <file> (exact YouTube 1280×720, under 2MB) and copy out with a clear name.

Mascot contract

The mascot must stay on-model every time — same colors, same outline weight, same face. That consistency is the brand recognition; a mascot that drifts is worth less than no mascot.

Costumes and props are fine (hat, vest, holding an object) as long as the body and face still read as the same character.

HARD SAFETY RULE — never violate: do NOT pose the mascot with a straight raised outstretched arm / open palm. With a crowd or army behind it, that reads as a fascist salute and will get the thumbnail flagged. Commander/hero poses must keep both arms bent and occupied (holding a banner or sword, hand on hip, fist raised, presenting to the side). Vary poses across a batch — don't reuse an identical stance.

Full contract, expression menu, and safe-pose list → references/character-and-safety.md.

Series rule

For a numbered series, never bake the episode number into the thumbnail — it makes viewers watch only Part 1. Give each episode a distinct, standalone hook that fits that episode, and vary the accent color and pose so the covers don't look like copies of each other. Copy-paste covers read as "already watched".

References

Requires the codex-fleet skill for the underlying Codex CLI image runner.