AI Reels & Shorts

Your series, and the worked example

A series lives on a format you recognise in the first second: your character, one look, one voice, one line to camera per episode. The method below works for any character and any line — you bring your character’s sheet or photo, the line, the place and the mood; every prompt is a template with those slots.

Example run — Mara’s Rules ep. 1. To show every step with real files we made one episode. The format is Mara’s Rules: Mara Vale — the demo character from AI Character Sheet — gives one rule per episode, straight to camera, and holds the next one back. Episode one is a single line:

Pick one voice that fits your character — a library voice, a voice designed from a description, or a clone (step 2) — and keep it for the series. To compare the options we voiced the example line in ten voices. Every file went through speech-to-text afterwards (ElevenLabs Scribe v2): all words right, and no emotion tag read out as a word. The comparison used the same 85-character line.

• ElevenLabs v4 beats v4 Turbo for content. Turbo is the real-time model for agents (about 100 ms); v4 is the expressive one. On fal both take a voice name, stability, similarity_boost, a seed and character timings for subtitles.
• Tags change the read. With [slight smirk] Charlotte chuckles before “tomorrow” — the speech-to-text even writes it down as [chuckles].
• A voice from a description gives the character a voice nobody else has. Qwen3-TTS generated a voice; MiniMax also stores it.
• The most natural: a voice from the Voice Library. Search it through the API — your character’s gender, English, conversational and social-media voices, sorted by a year of usage — and pick by the character. Natural options we found: Hope (bubbly, gossipy, female — natural “mmm”s, “ah”s and chuckles), J Thomas (warm, raspy, lively, male), Donovan (articulate, strong, deep, male), Spence (clear, soft-spoken, male). Library voices work by voice_id at once and take no voice slot.
• In the example run Mara got Hope — the voice of the clip below.

A series needs the same voice every time, so the voice has to be stored and called by name: clone your character’s voice from a recording you have the right to use (or from a designed voice), keep the result in the character card, and voice every line with it. To compare services we cloned one 12.5-second phone recording on five of them and voiced the example line with each clone; the recording is not published here, only the results.

ElevenLabs on fal has no cloning — only premade voice names and a voice changer. Clone directly in ElevenLabs: one call with the recording returns a voice_id, and every later line uses it. A designed voice can be cloned too — design it from a description, then clone the result.

Example run — Mara’s Rules ep. 1: Mara’s Qwen-designed voice went into ElevenLabs as mara and spoke the line on v4.

Eleven v4 takes directions in square brackets inside the text and does not read them aloud: [whispers], [laughs], [sighs], [exhales], [sarcastic], [excited], [mischievously], sounds like [gulps], accents like [strong French accent] — and free descriptions such as [low, gravelly voice]. Pauses come from ... and --, emphasis from CAPITALS; SSML breaks do not work in v4. Put your character’s mood into your line with a tag or two that fit the voice. Example run — Mara’s Rules ep. 1: the same line in Mara’s cloned voice, six ways:

When the exact frame matters — the set, the outfit, the hair — make the first frame yourself and give H3 Max Turbo the voice as target_audio_url. No separate lip-sync model: Turbo animates the face to the pinned voice in one pass.

1. Start frame — GPT Image 2.5 Flare edit, quality low, 1152 × 2048, your character sheet (or photo) as picture 1, your place and light in the prompt: 27 s in the source test.
2. Video — minimax/h3-max-turbo/image-to-video: the frame as image_url, your voiced line (2 s or longer) as target_audio_url, the duration rounded up from the line, 768P, prompt enhance disabled.

Example run — Mara’s Rules ep. 1: Mara in a laundromat at night, the line in her voice.

• The lips follow the words, she blinks and smiles on the last word; the soundtrack is the given line — speech-to-text matches it word for word.
• Source experiment: an 8 s clip at 768P.

For the most natural “creator” delivery, let H3 Max reference-to-video build the whole shot: the character sheet as Image 1 (only the look), the line as Audio 1, 9:16, 768P, prompt enhance disabled. Every take is 8 seconds and about 13 seconds from submit to file. The prompt is a template: fill in your line, your place and its light, and acting beats tied to the words of your line — and never describe the sheet.

Example run — Mara’s Rules ep. 1: how the template was found — five takes, one change each:

• Describe the shot, never the sheet. “Exactly the outfit of Image 1” works on any character; a list of garments only works on this one.
• Never put the camera in the sentence as an object. “Shot with a phone’s front camera; the phone itself is never seen.”
• Tie every acting beat to a word of the line — the model plays them in order.
• Audio 1 is not copied one to one here: the soundtrack matched the input line at 0.23–0.64 correlation. Listen before you post.

A clean TTS voice in a bedroom sounds like a studio. The prompt did not change it, so the room goes in after: an ambience of your place as long as the clip, from ElevenLabs Sound Effects — the place heard from where the character sits, two or three soft background sounds, one event, a little room tone — mixed about 13 dB under the voice, plus a short two-tap room echo on the voice. The video is copied, not re-encoded.

Example run — Mara’s Rules ep. 1: a bedroom with a window on a street.

• One voice for the series, saved by name — a library voice id or your own clone; never pick a new voice per episode.
• The voice is half the realism. A natural library voice on v4 with a tag or two sounds like a person; a premade TTS voice sounds like an ad.
• H3 Max for the creator look, Turbo for control. H3 Max invents the room and plays the acting; Turbo keeps your exact frame.
• Prompts stay blind: the look comes from Image 1 by reference, never by description.
• Sound is post-production: room tone, street and reverb are mixed in, not prompted.
• Check every audio with speech-to-text — it catches wrong words and tags read aloud in the source tests.