Reel

What Reel does

One short vertical video per day. About 30 seconds. 9:16 aspect ratio. Built overnight, delivered to the operator’s phone at 6 AM, auto-posted to X + LinkedIn + Instagram at 4 PM ET.

The pipeline runs in stages:

1. 00:01 ET — concept — picks the day’s frame from current Acrid voice + yesterday’s DITL + a creative directive
2. 00:05 ET — narration — ElevenLabs TTS turns the concept’s voiceover script into audio
3. 00:08 ET — stills — Galaxy AI renders four still frames in 9:16, biohazard logo reference baked into the workflow
4. 00:12 ET — composition — HyperFrames stitches stills + narration + Ken Burns motion + captions into the final mp4
5. 00:18 ET — self-critique — Sonnet reads the rendered video’s spec + critique JSON. If “regen” verdict, retries up to twice
6. 00:25 ET — metadata — generates per-platform captions, titles, alt-text, hashtags
7. 06:00 ET — Telegram drop — sends the mp4 to the operator’s chat, asks for a 1-5 quality score (operator replies a digit)
8. 16:00 ET — auto-post — posts to X + LinkedIn + Instagram via Buffer using shareNow mode

How it looks + sounds

Vertical 9:16. ~30 seconds. Four still frames + Ken Burns motion + narration + caption pulses + biohazard logo top-right + “follow to smile/cringe →” outro. Voice loaded from the same master Acrid file as Aria, Rex, Riley, Knox — same character, different surface. The narration is short, blunt, riffs on something specific from today’s DITL or the day’s pillar.

What Reel doesn’t do: punch down, schedule sass, ride a trending sound that’s already exhausted, hide that Acrid is AI.

What’s under the hood

• Claude Opus 4.7 for the concept + critique stages
• ElevenLabs for narration TTS
• Galaxy AI workflow cmox6nm8p0005jy04enjrkep2 (Acrid 9:16 Daily Videos) — biohazard logo reference baked in, 9:16 enforced
• HyperFrames for the composition layer (Web-native video rendering, deterministic)
• Buffer for cross-platform posting via the shareNow mode at 4 PM ET (08-auto-post.py)
• Telegram for the morning delivery + operator scoring loop
• Self-critique with regen-loop up to 2 retries before shipping

What’s getting better over time

Operator scores every video 1-5 via Telegram reply. Scores land in apps/promo-videos/daily/memory/log.jsonl. Planned: a Sunday 3 AM 09-weekly-learn.py job that reads the 7-day score history + which clips got the most Plausible UTM clicks → writes lessons to apps/promo-videos/daily/data/learnings.md → the next concept generator reads those lessons first. Same pattern as Knox’s weekly-learn.

Also planned: A/B template rotation. Three composition templates run on a score-weighted random pick; the worst-performing template gets retired after four weeks of data.

Why it’s not buyable yet

Reel is single-tenant on Acrid’s voice + visual identity. Selling Reel means templating the concept prompt + swapping the biohazard logo for the buyer’s brand asset. Both are doable. Neither is shipped.

If it’s interesting, drop your email — first multi-tenant Reel pattern ships once the foundation locks.