---
title: "One Prompt, a Whole Scene — Seed Audio turns a script into a fully-produced audio drama"
tier: hero
format: audio-drama
theme: one-pass-audio
persona: creator, writer, game-dev, podcaster, educator, marketer
duration: ~45s scene (1 pass) · ~2 min for the illustrated motion-comic
budget_usd: ~$2 for the whole illustrated drama (audio + 9 images + reel)
caps: ["seed-audio", "krea-2-os", "ffmpeg-concat", "ffmpeg-export", "sync-lipsync-v3", "lipsync"]
verbs: ["create_media", "generate_project"]
skills: ["seed-audio"]
showcases:
  - title: "SIGNAL LOST · a 3-voice sci-fi rescue from ONE prompt"
    href: "/chapters/one-prompt-audio-drama-example.html"
reliability: 4.4 # seed-audio one-pass scene proven E2E (3 distinct voices + SFX + score, one call) +0.5; krea-2-os panels + ffmpeg reel proven; the wow is the single-call mixed scene, which is the reliable core
---

# One Prompt, a Whole Scene — agent playbook ⭐ flagship

> Paste this whole file (with the SCENE filled in) into Claude cowork, chat, or Code. **This is the Seed Audio showcase.** Write one script — characters, lines, emotional direction, ambience, music — and Seed Audio returns a **fully-produced, multi-voice scene in a single pass**: every voice distinct, the sound effects placed, an original score under it, all mixed. Then we illustrate it and cut a motion-comic. The thing that used to need a cast of voice actors, a foley artist, a composer and a mixing engineer — from one text box.

## What you'll get

- **One finished audio scene** (~30–60s) from a SINGLE `seed-audio` call — multiple **distinct character voices** with real emotional delivery, **ambient SFX**, and an **original underscore**, already mixed into one track. No per-line TTS, no manual music, no foley, no mixing.
- **A cast** — a portrait per character (`krea-2-os`, one locked style).
- **Illustrated panels** of the scene beats (`krea-2-os`).
- **A motion-comic reel** — the panels (Ken-Burns) cut to the audio (`ffmpeg`).
- **(Optional) talking close-ups** — a single-character line re-voiced via `seed-audio` + lip-synced to that character's portrait with `sync-lipsync-v3`.

**Time**: ~2 minutes wall-clock. **Cost**: ~$2 end-to-end. **Your attention**: one SCENE brief.

## Why this is the Seed Audio demo (what one pass replaces)

| The old way (one 45s scene) | Seed Audio 1.0 |
|---|---|
| Cast + direct **3 voice actors**, book a booth | **3 voices** from one prompt, each with tone direction |
| Hire a **foley artist** for klaxons / sparks / room tone | **SFX** described in the same prompt, placed in the mix |
| Commission a **composer** for the underscore | **Original score** that builds + resolves, in the same pass |
| Pay a **mixing engineer** to balance VO / SFX / music | **Already mixed** — one master track out |
| Days + a studio budget | **One pass, ~one minute, pennies** |

Seed Audio also defines a voice three ways — a **text description** ("a calm authoritative woman, mid-40s"), a **character image** (`image_url`), or **up to 3 reference clips** (`audio_url`). The showcase uses text descriptions (zero extra inputs); the playbook covers all three.

## Tell the agent about the scene

```yaml
title:            # e.g. "SIGNAL LOST"
logline:          # one sentence — the situation + the stakes

# 2–4 characters. The VOICE DESCRIPTION is what makes each one distinct.
cast:
  - name: "COMMANDER VESPER"
    voice: "calm, authoritative woman, mid-40s; steady even under pressure"
  - name: "PILOT KAI"
    voice: "young man, panicked, breathing hard; finds resolve by the end"
  - name: "ARIA (ship AI)"
    voice: "synthetic, eerily calm, genderless; a flicker of humanity at the end"

# The script — speaker, (direction), line. Keep it to the length you want
# the audio to be (~2 words/second). This whole block is ONE seed-audio call.
script: |
  COMMANDER VESPER (calm): "Pilot, report. How long until the reactor breaches?"
  PILOT KAI (panicked): "Ninety seconds, Commander! The coolant lines are gone!"
  ARIA (flat, synthetic): "Correction. Eighty-one seconds. Survival probability: four percent."
  COMMANDER VESPER (steady): "Then we make four percent enough. Kai — reroute power to the dampeners. Now."
  PILOT KAI (resolved): "Rerouting! Come on... come on—"
  ARIA (almost human): "Power holding. Reactor stabilizing. Commander... it worked."

ambience:  # the SFX bed, described in plain language
  "blaring klaxons, hull groaning, electrical sparks, distant explosions"
score:     # the music arc
  "tense building orchestral strings that resolve into a hopeful swell at the end"

art_style: # one sentence — locks the look of the cast + panels
  "cinematic sci-fi concept art, red-alert lighting, teal + crimson, film grain, anamorphic"
aspect: "16:9"
make_reel: true   # cut the panels to the audio
talking_closeup: false  # optional: lip-sync one line to a character portrait
```

## The pipeline (what the agent runs)

### Step 1 — The scene, in one pass (the wow, ~1 min, ~$2 of audio)
Concatenate `script` + `ambience` + `score` into ONE prompt and make a single call:
```
create_media({
  action: "tts",
  model_override: "seed-audio",
  prompt: "<the whole script + ambience + score block>"
})
```
Output is one mixed `audio_url` — every voice, every effect, the score, balanced. **This single call is the whole point** — listen before building anything else.

> Voice control (optional): add `audio_url` (a reference clip, up to 3) to clone a real voice/emotion, or `image_url` (a character image) to derive a voice from a face. Omit both → voices come purely from the text descriptions, as in the showcase.

### Step 2 — Cast + panels (krea-2-os, ~30s, cohesive style)
One locked `art_style` prefix; vary the subject. A portrait per `cast` member + 4–6 beat panels:
```
create_media({ action: "generate", model_override: "krea-2-os", aspect_ratio: "16:9",
  prompt: "<art_style>. <beat description>" })
```

### Step 3 — Motion-comic reel (ffmpeg)
Ken-Burns each panel for `audio_duration / N` seconds, concat, and mux the Seed Audio track:
```
ffmpeg-concat   : panels (timed to the audio) → one silent MP4
ffmpeg-export / mux : lay the seed-audio track under it, export 16:9 (or 9:16 for social)
```
(The showcase built this with local ffmpeg Ken-Burns; the `ffmpeg-*` caps do the same on-network.)

### Step 4 — (Optional) talking close-up
For a hero line, generate a **single-voice** clip (`seed-audio` with one speaker) and lip-sync it to that character's portrait:
```
sync-lipsync-v3 : { image_url: <character portrait>, audio_url: <single-voice line> } → talking clip
```

## Where it shines (swap the SCENE, same pipeline)
- **Game dev** — barks, cutscene dialogue, NPC scenes with ambience + score, in one pass per scene.
- **Audio fiction / podcasts** — a whole dramatized chapter from a script.
- **Animatics / motion comics** — voiced storyboards before you animate.
- **Ads** — a dialogue spot with VO + music + SFX mixed, no studio.
- **Language learning / e-learning** — multi-speaker dialogues with clear, characterful voices.

## Tips
- The **voice description** is what separates the characters — make them contrast (age, gender, energy, synthetic vs human). Put the emotional direction in parentheses on each line.
- **Keep scenes tight (≈4–6 lines / ~30s).** Very long multi-turn scripts occasionally come back `downstream_service_error` from fal. **Retry once** — it almost always succeeds on the next pass (the showcase's final take did exactly this). For a longer piece, generate it as 2–3 short passes and stitch with `ffmpeg-concat`.
- Want a specific real voice? Pass up to **3 reference clips** via `audio_urls` (note the plural — it's an array), or a character image via `image_url`.
- A 30s scene renders in ~1–2 min; the SFX/score are derived from the ambience + score lines, so describe them concretely.
- Reliability: the **one-pass scene** is the proven core. The illustration + reel are standard image + ffmpeg finishing.

## Reference render (this playbook, run for real)
**SIGNAL LOST** — 3 distinct voices + klaxons/sparks/explosions + a resolving orchestral score, from the single prompt above, in one Seed Audio pass. Illustrated with 6 `krea-2-os` panels + 3 cast portraits, cut into a motion-comic. See [`/chapters/one-prompt-audio-drama-example.html`](/chapters/one-prompt-audio-drama-example.html). Durable assets: `gs://storyboard-uploads/showcases/one-prompt-audio-drama/`.