Explorations Create yours

Scenario Exploration · A Cast of One

Four characters. One actor. No suit.

A Cast of One is a Scenario exploration: a pilot preview whose performer is a synthetic stand-in made with Scenario, not a real person. Characters, names and places are invented; the stand-in's voice was generated with her take, and each character speaks in a preset voice.

Phone takesynthetic stand-innot a real person
The knightSer Corran Vale

One person plays every role of a tavern scene, filmed on a phone, one role at a time. Scenario turns each take into a different character with its own body and voice. Drag the line across the picture to see who was really there.

The cast

Pick a role: Knight · Goblin · Wizard · Raven. Its card shows the body, the voice and what held. The film is the pilot's 20 s preview cut.

20 s cut of the pilot: one line, four voices, then the reveal. The reveal at the top of the page has the newer knight and wizard, and the raven speaking. Made with: Take Minimax H3Knight P-Video Replace + Sync-3 LipsyncGoblin, wizard Kling V3 Pro Motion ControlVoices ElevenLabs Voice ChangerRaven GPT Image 2.5 Sunburst

The scene

The Crooked Lantern, a roadside tavern at night. A knight, a goblin merchant and a wizard share a table. Above them, on the rafter, a raven that never blinks.

    The pilot filmed one line (S4) and cast it four ways, to test every method on the same performance. The other lines wait for the real shoot.

    What held, and what did not

    The pilot pushed one take through every method the concept named. Two comparisons decided which version each character keeps.

    The Seedance 2.5 testthree comparisons

    On 27 September each role was re-made with Seedance 2.5 from the same take, with the character sheet as a reference, and judged the same way: the face, the performance kept, the mouth at the key words. It held on the knight and the wizard, which are now on Seedance; the goblin stays on Kling. Both goblin takes were re-made later that day from a still whose lips cover his teeth.

    How a take becomes a character

    Two lanes run in parallel: the voice and the body. The mouth pass, when it is needed, waits for both.

    Measured time 0 s 0 CU

    Replay of the pilot's measured run: every duration and credit (CU) figure is the job's own record on Scenario, played 20 times faster.

    The shoot

    • One phone, one tripod. Locked off, horizontal, 24 fps, exposure and focus locked.
    • Same frame for every role. Mid-chest up, a plain wall behind, tape on the floor.
    • Soft, even light. Two panels either side of the lens, nothing behind.
    • Hands down, face clear. Turns to three-quarter, never full profile.
    • Eyeline marks. One tape cross per character, one for the raven on the rafter.
    • 10 to 15 s a take. Inside every model's limits, with a breath at each end.

    The agent skill and the making-of

    The agent skill packs this pilot's method for your own take: one character sheet per role, the body transfer that held for each type of character, a new voice that keeps the actor's timing, the checks, and a reveal pair. The making-of follows the pilot from the stand-in's take to the reveal.

    Use the agent skill

    1. Unzip it into your agent's skills folder so that scenario-one-actor-many-characters/SKILL.md sits inside: ~/.claude/skills/ for Claude Code, ~/.codex/skills/ for Codex.
    2. Connect Scenario MCP to the same agent.
    3. Film one line on a phone on a tripod (your own face, or someone who agreed in writing), then start with this prompt:

    Use the scenario-one-actor-many-characters skill: turn my phone take take.mp4 into two characters of my own, each with its own look and voice, and make a reveal pair of the take and each character on the same frames. Ask me for the characters first, and show me the character sheets and the price of each video before you run it.

    How it was made

    One take, four characters, 69 jobs

    One headless Claude Opus 5.5 session ran the pilot through Scenario MCP in four runs: the pilot on 24 September, the page and a shoot kit on 25 September with no generation, a Seedance 2.5 test on 27 September, and a revision the same day that redrew the goblin's mouth and made the raven speak. It pushed one 8 s line, performed by a synthetic stand-in, through every method on the same take: 69 jobs and 10,725 CU, checks included, every call priced with a dry run first.

    The take

    GPT Image 2.5 Sunburst painted the first frame as a phone self-tape; Minimax H3 animated it into the 8 s take with native audio.

    The cast

    GPT Image 2.5 Sunburst painted the four character sheets and the tavern in one style, then redrew the goblin's lips and closed the raven's beak on its still.

    The voices

    ElevenLabs Voice Isolator cleaned the take's voice, ElevenLabs Voice Changer turned it into one preset voice per character, the raven included, and Gemini 3.5 Transcribe checked the words before and after. Minimax H3 Max Lip Sync moved the raven's beak to its voice.

    Each role, as measured

    What held

    Agent skill · ZIP · 21 KB

    Use the agent skill

    Download the agent skill (ZIP, 21 KB)