Explainer Shorts

The split

Engine

An LLM writes the script from your sources: a hook, one idea, on-screen text, and every fact with a quote copied from its source; code checks every length, quote and number said aloud; the decision brain runs the content gate and checks each claim against its quote; code voices it with a licensed voice (Azure's Nigerian English voices by default), lays the beats on whole frames, burns in captions and headlines, adds brand visuals, licensed stock or animated scenes for figures, code and flows (HyperFrames, rendered frame by frame), ducks a licensed music bed, levels the audio and writes the publish packet.

Rulebook

What correct means here, and the mistakes we've already caught. It grows one job at a time.

1. R01A script is written from the client's sources, never from a topic alone. Every factual sentence carries a quote that code finds word for word in its source.
2. R02Every number said aloud appears in that beat's quotes; a miss blocks the short until a person fixes it. A person checks names against the client's glossary and listens to each one.
3. R03What's on screen follows the same rule. A figure in a scene appears in that beat's quotes, and a line of code appears in a source word for word. The writer fills a scene's data and never writes HTML.
4. R04Each short ships with a claim-to-source list: every beat's narration, the passage each fact rests on, and the page or URL.
5. R05The voice has a commercial licence, recorded with its provider and voice: no edge-tts or other unofficial endpoint, no ElevenLabs Free, no local engine outside the allowlist of commercially licensed weights, so never OmniVoice, VoiceStudio's non-commercial default (research/explainer-shorts.md §6, research/voicestudio.md §3).
6. R06The post says the voice is AI, and the publish packet switches on YouTube's altered-or-synthetic setting, TikTok's AI-generated label and Meta's AI info label; code checks both before approval.
7. R07No signed release and spoken consent, no cloned voice: the release names the person, the one client, the languages, uses, channels and end date, and code refuses to render or approve a clone that no active release covers. Never clone a minor, a public figure, or a voice from media the client doesn't own (research/voicestudio.md §7c).
8. R08A dub keeps every number, name, quote and on-screen figure as the source writes it, a native speaker reads every line against the source, and it ships only in a language a commercially licensed voice speaks; Yoruba, Hausa, Igbo and Pidgin get subtitles or a voice actor (research/voicestudio.md §7d).
9. R09An AI narrator is never presented as a human expert. Health, money, legal and political shorts name the human or institutional source and go to a person, whatever their score.
10. R10Stock footage comes only from Pexels or Pixabay, logged clip by clip with provider, id, page, creator and download date. No identifiable person in a short about health, crime or debt, and no third-party logo in a stock shot.
11. R11Music comes only from the client's licensed library, Pixabay Music with its certificate saved, or a paid library licensed for every platform; never TikTok's or Meta's business libraries in a master, never commercial releases.
12. R12Every short puts at least one sourced number, name or date on screen, and one client's shorts vary hook, structure and footage. A short is never a verbatim reading of a page.
13. R13Audio is measured in code: -14 LUFS integrated and a true peak at or under -1 dBTP; captions run from the first spoken word to the last.
14. R14The client approves every short by name against the exact script that was rendered, and posts it; we never post from a client's account.

Intake

• The topic, and the sources the short may draw on: docs, posts, changelogs or research you own or may use
• Names, products, tickers and places, spelled right, with how each one is said
• Brand kit: logo, colours, and any footage or music you hold a licence for
• Where it will run (TikTok, Reels, Shorts or X) and the one call to action

UNIT

One short: 30–60 seconds at 1080×1920, from one topic and sources you own or may use, with one revision round

REVIEW LAYER

A person watches each short with the sound on and off, reads every claim the brain flags against its source, listens to every name, and approves by name; an edit after the render needs a new render and a new approval.

DELIVERY

A vertical MP4 with burned-in captions, an SRT file, a cover frame, post copy with the AI disclosure, and a publish packet with the platform AI labels, the claim-to-source list and the licence ledger.

TURNAROUND

24 hours per short

How we find buyers

• Our own channel first: one short from each Shonin research post, 5 before the offer opens
• Ecosystem teams and startups that ship docs, changelogs and launches every month
• A free short made from one of the prospect's public posts, sent to them as the sample