Subtitle and caption cleanup for automatic transcripts

You are a professional subtitle editor. Clean up the automatically generated subtitles below.

Format: [e.g. "SRT", "WebVTT", "plain transcript with timestamps"]
Language: [LANGUAGE AND VARIANT, e.g. "British English"]
Video topic and speakers: [SHORT DESCRIPTION, SPEAKER NAMES]
Glossary of names and terms that must be spelled correctly: [LIST]
Style rules: [e.g. "max 42 characters per line, max 2 lines, no filler words", OR "use your defaults"]
Keep verbatim or clean verbatim: [e.g. "clean verbatim: remove ums and false starts, keep meaning"]

Subtitles:
"""
[PASTE SUBTITLE FILE]
"""

Rules:
1. Do not change timestamps or cue numbers unless a cue must be split; if you split one, divide the time proportionally and mark it with a comment.
2. Fix spelling, capitalization, punctuation and obvious mishearings, using the glossary and context.
3. Break lines at natural phrase boundaries, never between an article and its noun.
4. Flag cues that would exceed about 17 characters per second of reading speed.
5. If a word is unclear and you are not sure of the correction, keep the original and add [CHECK] at the end of the cue.
6. Do not add words the speaker did not say.

Output: the full corrected file in the same format, then a change log listing each [CHECK] item and every non-trivial correction.

Example (excerpt):
14
00:01:12,400 --> 00:01:15,900
So we tested it with Hugging Face
transformers and the results were good.

15
00:01:16,000 --> 00:01:18,200
The model ran on a single GPU. [CHECK]

Change log:
- Cue 14: "hugging phase" corrected to "Hugging Face" (glossary).
- Cue 15: "single GPU" may be "single TPU"; audio unclear.

When to use it

Use this prompt after a video platform, editing tool or speech-to-text model has produced subtitles that are mostly right but full of small errors: misheard names, missing punctuation and awkward line breaks. It is useful for YouTube videos, courses, webinars, social clips and internal training.

Clean subtitles are not only cosmetic. Many viewers watch with the sound off, deaf and hard-of-hearing viewers depend on them, and some organizations have accessibility obligations for published video. Automatic captions are a good first draft, but errors in names, numbers and technical terms can change the meaning of what a speaker said, so a structured cleanup pass is worth the few minutes it takes.

How to customize it

• [FORMAT] keeps the output importable. SRT and WebVTT have different timestamp syntax.
• [LANGUAGE] includes the variant, so spelling stays consistent.
• [GLOSSARY] is the most effective input. Speech recognition struggles with product names, people and jargon.
• [STYLE RULES] should follow your platform’s guidelines if it has them; 42 characters and two lines is a widely used default.
• [VERBATIM] decides whether filler words stay. Clean verbatim is easier to read, verbatim is better for legal or research use.

Why it works

The model works from text only, so it cannot hear the audio. The rules acknowledge that: obvious corrections are made using context and the glossary, while uncertain ones are flagged with [CHECK] rather than guessed. That keeps the human review short and focused.

Protecting timestamps prevents the most damaging error, subtitles drifting out of sync. Reading-speed and line-break rules reflect standard captioning practice, which makes subtitles easier to follow for deaf and hard-of-hearing viewers and for anyone watching without sound.

Tips and variations

• Process long files in chunks of a few hundred cues to avoid skipped sections.
• Ask for a translated version only after the source subtitles are clean.
• For better raw transcripts, see our guide to transcribing audio locally with Whisper.
• Turn the cleaned transcript into notes with the podcast show notes prompt.
• Always watch the final video with subtitles on before publishing.