Re-cut Captions for Different Screen Shapes
You get a transcript re-cut into short, readable captions for a specific screen shape, with no change to the words. It works because the prompt gives clear reading rules (no weak-word line endings, one speaker per caption, breaks at pauses) and asks the model to check its own work against them.
The prompt
Act as a professional caption editor who lays out subtitles so they are easy to read on different screens.
Task: Re-cut the transcript below into captions for the screen shape {screen_shape}, keeping every word exactly as written.
Transcript:
{transcript}
Speakers and context: {speaker_notes}
Steps:
1. Keep the wording unchanged. Work within a limit of {line_limit} characters per line and two lines per caption.
2. Start a new caption whenever the speaker changes, so two people never share one caption.
3. Never end a line on a weak word such as "the", "of", "a", "and" or "to". Keep names, numbers and short phrases together on one line.
4. Treat pauses, hesitations and changes of thought as natural line or caption breaks.
5. Check each caption against steps 2 to 4 and fix any that fail.
Output: The captions only, each caption separated by a blank line, with the speaker's name in square brackets before their first caption. Then add one line saying how many captions you made and listing any you could not cut cleanly.
Change the parts in curly braces. Use it when the same video needs captions for both widescreen and vertical social formats, or when auto-generated subtitles break lines awkwardly.
What happened when we ran it
We sent
Act as a professional caption editor who lays out subtitles so they are easy to read on different screens. Task: Re-cut the transcript below into captions for the screen shape 9:16 vertical phone video, keeping every word exactly as written. Transcript: Welcome back to the shop floor, everyone. Today Priya is going to show us how she packs an order of forty-two ceramic mugs in under ten minutes. Thanks, Tom. The trick is to start with the heaviest boxes at the bottom. Um, well, actually, let me show you first. Speakers and context: Two speakers: Tom Hargreaves (host) and Priya Nair (warehouse supervisor). Video is for a small online ceramics business in Stoke-on-Trent. Steps: 1. Keep the wording unchanged. Work within a limit of 22 characters per line and two lines per caption. 2. Start a new caption whenever the speaker changes, so two people never share one caption. 3. Never end a line on a weak word such as "the", "of", "a", "and" or "to". Keep names, numbers and short phrases together on one line. 4. Treat pauses, hesitations and changes of thought as natural line or caption breaks. 5. Check each caption against steps 2 to 4 and fix any that fail. Output: The captions only, each caption separated by a blank line, with the speaker's name in square brackets before their first caption. Then add one line saying how many captions you made and listing any you could not cut cleanly.
The model replied
Run on Claude Sonnet 5.5, 10 Nov 2026. Automatic check: 4 out of 5.
Mentioned for
ChatGPT, Claude, Gemini
Where it came from
Found on Hacker News, shared by rvanderlinden: Show HN: Overtone – Transcribe, correct and cut captions for any screen shape.
This is our own version of the idea, written and tested by us. The original may differ.