---
url: https://superduperai.co/library/reaction-duet-style-format/
title: Reaction / duet style: respond to a clip in 30 seconds without stealing it
description: The split-screen or picture-in-picture response format, written so the reaction adds a fact or a demonstration instead of a face. The prompt plans the pause points and what you say at each.
category: formats
platforms: [tiktok, reels, shorts]
works_with: [claude-code, claude, chatgpt, any-llm]
tags: [reaction, duet, stitch, tiktok, reels, reach, script]
difficulty: intermediate
verified: false
created: 2026-11-01
---

# Reaction / duet style: respond to a clip in 30 seconds without stealing it

The split-screen or picture-in-picture response format, written so the reaction adds a fact or a demonstration instead of a face. The prompt plans the pause points and what you say at each.

## Prompt

```text
Write a reaction / duet script in which I respond to a clip about {topic}. The format works because the source clip supplies the hook and the pacing, and the viewer stays to see whether I agree. A reaction that only nods is skipped; each pause must add something the source did not have (a correction, a number, a demonstration, a missing step).

Input: the source clip transcript with timestamps, and my position (agree / disagree / 'yes, but').

Structure (30 s, 9:16, source clip on top or left, me below or right):
- 0-2 s: let the source play its hook. My on-screen text over my half: my stance in 3 words ('this is wrong', 'half right', 'do this instead').
- Then 3 PAUSE POINTS. At each: freeze or lower the source, I speak for 5-7 s, one point each, max 18 words. If I can show it instead of saying it, the shot note says what I hold up or do.
- Between pause points the source plays 3-4 s at full volume.
- 27-30 s CLOSE: one sentence verdict. No download ask; if my demonstration used a product, it was visible, that is enough. The comment ask is 'who's right?'.

Word budget: 55-65 spoken words at 165 wpm (I am speaking for about 20 of the 30 seconds). On-screen text: the stance at 0-2 s, then one 3-word label per pause point.

Return: a table (time, what the source is saying, what I do, what I say, on-screen text), and a 1-line note on which of my three points is the strongest, so I can lead with it if the source clip is shorter than planned.

Source transcript: {transcript with timestamps}
My position: {agree / disagree / partly, and why}
```

## Example

Source: a 25-second clip claiming you should shoot product photos at night with a ring light. My position: disagree, window light is better and free.

| Time | Source says | I do | I say | On-screen |
|---|---|---|---|---|
| 0-2 s | 'Stop shooting in daylight, here's why' | Listen, eyebrows up | (nothing) | DO THIS INSTEAD |
| 2-8 s | 'ring light gives you control' | Hold up the mug, lamp on, then off | Control, yes. Also a flat face and a white ring in every reflection. Look. | RING REFLECTION |
| 8-11 s | 'daylight changes every minute' | Let it play | | |
| 11-18 s | 'so you can't match shots' | Turn the mug by the window, side on | It changes slowly. Side on at the window, shoot all your products in one twenty-minute go. Done. | 20 MIN, ONE GO |
| 18-21 s | 'and you need it to look professional' | Let it play | | |
| 21-27 s | 'that's what buyers expect' | Show two photos side by side on the phone | Window on the left, ring on the right. Which one would you buy? Same mug. | SAME MUG |
| 27-30 s | (ends) | Face | Free light, five steps from the kitchen table. Who's right? | |

Strongest point: the side-by-side at 21-27 s. Lead with it if the source is short.

## Tips

- Only react to a clip you can add a demonstration to. 'I disagree' with a face is not content; 'I disagree, look' is.
- Keep the source audible between your points. Muting it for the whole video turns a duet into a lecture with a stranger on screen.

## Works with

- Claude Code: Runs as a skill or a prompt in the terminal; can read your files and write the result.
- Claude: Paste into claude.ai or the Claude app.
- ChatGPT: Paste into ChatGPT; works with any GPT model.
- Any LLM: Model-agnostic: Gemini, Mistral, a local model.