ChatGPT Prompts to Write an Audio Drama Script (That Actually Work by Ear)

The best ChatGPT prompts for an audio drama script tell the model three things up front: that the script is for the ear only, that every line needs a speaker tag, and that sound cues go on their own lines. Without those rules you get a screenplay with camera directions, or a short story with quotation marks. Below are five prompts you can copy, the fixes for what the model usually gets wrong, and how to get the finished script into multi-voice audio.

One note before the prompts: these are for your own use in ChatGPT or any other chat assistant. You still decide the story, the characters and what stays in the final draft. The prompts just make the model produce something shaped like an audio script instead of something you have to reformat line by line.

Why a plain "write me an audio drama" prompt goes wrong

Ask a chat model for "an audio drama about a lighthouse keeper" and it will usually fall back on the two formats it has seen most: the screenplay and the prose story. The screenplay version is full of things a listener can never receive, like "INT. LIGHTHOUSE - NIGHT" and "She glances at the door." The prose version buries the dialogue inside paragraphs of description that a narrator would have to read aloud.

An audio script is different. Everything the audience learns has to arrive through dialogue, sound or a narrator line. The audio drama script format guide walks through the page layout in detail; the point here is that you have to state those rules in the prompt, because the model will not assume them.

Five prompts that produce ear-first scripts

Copy these as a starting point and swap in your own premise. Each one builds on the rules in the first.

1. The setup prompt. Use this at the start of every new chat so the rules stay in context:

You are helping me write an audio drama script. It will be heard, never seen. Rules: every spoken line starts with the character name in capitals and a colon. Sound cues go on their own line in square brackets, starting with SFX:. No camera directions, no scene headings, no descriptions of what characters look like or do silently. If the listener needs to know something, a character says it or a sound shows it. Confirm the rules before we start.

2. The scene prompt. Ask for one scene at a time rather than a whole episode:

Write scene one. Setting: a lighthouse during a storm, told only through sound. Characters: MARA (the keeper, 60s, clipped and dry) and ELI (a stranded sailor, 20s, talks too much). Goal of the scene: Eli asks to stay the night and Mara refuses, then changes her mind for a reason she will not say. Keep it under 500 words. Two characters only.

3. The voice-contrast prompt. Run this before any dialogue gets long:

Before writing more, give each character three speech habits a listener could recognize with their eyes closed: sentence length, favorite words, and one verbal tic. Make MARA and ELI as different from each other as possible.

4. The sound-first prompt. Useful for openings and transitions:

Open the next scene with sound cues only for the first three lines, so the listener knows where we are before anyone speaks. Then have the first line of dialogue confirm the place without naming it directly.

5. The table-read prompt. Use it on your own edited draft:

Here is my revised scene. Read it as a listener with no script in front of them. List every moment where you could not tell who was speaking, where you were, or what just happened. Do not rewrite anything; just list the problems.

The last prompt matters because it keeps the model in the role of reader rather than author. You fix the problems yourself, which keeps the voice of the script yours.

Fixing what the model gets wrong for audio

Even with good prompts, a few problems come back again and again. They are easy to spot once you know them.

• Characters who sound alike. Models tend to give everyone the same fluent, polite register. Go back to prompt 3, then cut any line that could belong to either character. If two voices still blur on the page they will blur in the audio, and the guide on what to do when two characters sound too similar covers the casting side of that fix.
• Exposition dialogue. Lines like "As you know, Mara, you have kept this light for thirty years" are the model trying to replace a camera. Ask it to move that information into conflict or a sound instead, or cut it and trust the listener.
• Silent actions left in. Watch for lines in brackets that describe a look or a gesture. If it has no sound, the listener cannot receive it. Turn it into a line of dialogue, a sound cue, or delete it.
• Too many voices per scene. Chat models like to add a third or fourth character to keep things moving. In audio, three voices in one scene is already a lot to track. How many voices an audio drama needs goes into where that ceiling sits.
• Names used too often, or never. Some drafts put a name in every line; others go a whole page without one. The audio drama dialogue guide explains how to place names in the first exchanges and then ease off.

Keep the model on one scene at a time

Long single prompts produce long, flat scripts. The model loses track of each character's habits, and scene three starts to sound like scene one. Working scene by scene keeps each piece short enough for you to read aloud, and gives you a natural point to edit before the next scene builds on the last.

A simple rhythm works well: setup prompt once, then for each scene a scene prompt, your own edit, and a table-read prompt on your edited version. Keep an outline in a separate document so you are steering the story rather than letting the chat drift. If the model introduces a new character or a plot turn you did not ask for, decide deliberately whether to keep it.

Clean the script before you produce it

When the scenes are done, copy only the script into a plain document. Leave behind your prompts, the model's replies about what it changed, and any Markdown symbols like asterisks and pound signs. The cleanup steps in turning a ChatGPT story into an audiobook apply here too, so there is no need to repeat them.

Then do one final check of the format: one speaker tag per line, the same spelling of each name every time, and sound cues on their own lines. Consistent tags matter more than they look, because they are what lets a tool assign each line to the right voice.

From chat to multi-voice audio

Once the script is clean, producing it is the quick part. In AudioProducer.ai you paste the script or import it as an EPUB, run Auto-Assign to tag each line by speaker, pick a voice for each character from the library, add sound effects, and download the result as an MP3. You wrote the script with your own prompts and edits; the tool handles the voicing and the render, and you decide where the finished audio goes.

Start small. Take the scene you like best, produce just that, and listen without the script in front of you. You will hear the problems a page hides, like two voices that are too close or a line that only made sense on screen, and you can take those notes back into your next round of prompts. Every other guide in this series is listed in all articles.

Your next step is one scene, not a whole season. Create a free account and hear your first scene; the free tier covers 1,200 words a month, which is enough to produce a full scene with two or three characters and hear whether your script works by ear.