Let a scheduled pipeline plan, script, fact-check, voice and post short Instagram videos: week one was about 2,100 views and zero sign-ups

Posting short videos every day takes a solo founder hours of scripting and editing, and it is hard to know which formats are worth repeating.

1. Every Monday, in the audience's time zone, a worker plans the week: up to two slots a day at the best times, each tagged core, sales or test, with a format, hook, ending and idea. 2. Core slots use Thompson sampling to pick formats, with no format allowed more than 60% of core slots over four weeks. 3. A Postgres cron job (pg_cron) checks every 10 minutes for approved slots due within 30 hours and queues them. 4. A small language model writes the script into a JSON scene file that follows a fixed schema (it never writes code), and a second model pass fact-checks the script. 5. A text-to-speech voice reads the lines and returns word timings, which set scene length and captions. 6. Remotion renders a 1080x1920, 30 fps video in a container that only runs when there is work, and the video is posted. 7. After seven days each video is scored on reach, watch time and actions against the account's own median, with caps on hits and flops; learning starts after five scored videos. The founder reviews a Monday report and fixes the system, not individual videos.

Week one across two new accounts (English and Italian): 16 videos, about 2,090 views, 57-58% swiped away in the first 3 seconds, about 6 seconds average watch time on 34-36 second videos, followers rose from 2 to 5 and 1 to 4, and zero sign-ups. The two sales videos did worse than average and nobody used the reply keyword. The best English video was one of the AI-made ones; the hand-made demo was the most skipped. Stated costs: about 0.1 cent per video for title and caption and about 1 cent for the fact-check.

Vendor: the author builds and sells the app being tested (HypeFox), so this is partly promotion. One week, 16 videos, two brand-new accounts, so the numbers say little yet; the learning step had no data in week one. The language and speech models are not named. Costs are the author's estimates and exclude rendering and hosting. No per-video human approval is described, which risks posting errors. Self-reported, no analytics screenshots checked here.

Instagram business accounts, a Postgres database with pg_cron, a text-to-speech service with word timings, Remotion, and the ability to write and host the worker code.

Keep a fixed share of test slots so the system keeps trying new formats, and wait for at least five scored videos before letting results change the plan.

Fill this JSON scene file for a 35-second vertical video about [topic] for [audience]: hook (max 8 words), 4 scenes with one line of voiceover each, an ending line and an on-screen caption per scene. Then list every factual claim in the script so I can check it.

• DEV Community (Mattia Maggioni, HypeFox) ↗ Vendor case study, published October 26, 2026

Source published: October 26, 2026. Last reviewed here: October 30, 2026. Spot a mistake? Tell us.