goulahai.com

Articles · Published · By Goulah

How to Make Accurate Captions for Your Videos With AI (Without Rewriting What People Said)

I'm a content creator and AI strategist, and I've been creating content for my own brands for about 9 years. Captions are one of those jobs that looks done the second the platform generates them, but a wrong word on screen is easy to miss and hard to un-see once a viewer catches it.

This guide shows a simple way to use AI to clean up captions without letting it rewrite what people said, then get them onto YouTube correctly. It uses YouTube's official help pages, the W3C's accessibility guidance on captions, and Anthropic's guidance for keeping AI answers grounded. The prompts work in any chat assistant that accepts long pasted text.

Why automatic captions need a human pass

YouTube can create captions automatically with speech recognition, and it says plainly that quality varies: automatic captions might misrepresent what was said because of mispronunciations, accents, dialects, or background noise. YouTube encourages creators to add professional captions first and to always review automatic captions and edit anything that wasn't transcribed properly. Source: Use automatic captioning (YouTube Help).

The W3C's Web Accessibility Initiative goes further. It says automatically generated captions don't meet user needs or accessibility requirements unless they're confirmed to be fully accurate, and that they usually need significant editing. Its example: the speaker says "Broil on high for 4 to 5 minutes. You should not preheat the oven," and the automatic caption reads "Broil on high for 45 minutes. You should know to preheat the oven." One wrong word flips the meaning. Source: Captions/Subtitles (W3C WAI).

The W3C also notes that captions are for people who are Deaf or hard of hearing, and for people who process written information better than audio. So "close enough" isn't a small problem for the viewers who rely on them.

1. Get a caption file to start from

You need a starting file with timestamps. If the video is on your YouTube channel, YouTube may already have automatic captions for it; it says they can take time to process and may not be ready right after upload. YouTube lists reasons they might not appear at all, including unsupported language, poor sound quality, a long silence at the start, or speakers talking over each other (source). How to download the caption track from YouTube Studio is in step 2 of How to Find the Best Short Clips in a Long Video With AI.

Download it as SubRip (.srt) if you can. YouTube recommends SubRip or SubViewer (.sbv) for people new to caption files, because they only need basic timing info and can be edited in any plain text editor. Here's what SubRip looks like, from YouTube's own example (source):

1
00:00:00,599 --> 00:00:04,160
>> ALICE: Hi, my name is Alice Miller and this is John Brown

2
00:00:04,160 --> 00:00:06,770
>> JOHN: and we're the owners of Miller Bakery.

Each caption has a number, a start and end time, and the text shown in that window. That structure is what you'll tell the AI not to touch.

2. Write down the words it will get wrong

Speech recognition struggles most with names and specialized terms, and the AI fixing the captions can't know how your guest spells their name unless you tell it. Before you prompt, make a short list: every person, company, product, show name, and piece of jargon in the video, spelled correctly. This list does most of the work.

3. Have the AI fix words, not rewrite them

The W3C's transcribing guidance is clear that captions should be accurate and honest: don't change, adapt, or add to what was said, it's usually not appropriate to correct grammar, don't censor, and if you can't understand something, mark it [unintelligible]. For most web content, it says it's acceptable to leave out non-substantive filler like "um" to make captions easier to read. Source: Transcribing Audio to Text (W3C WAI).

Those rules go straight into the prompt. Anthropic's guide on reducing hallucinations also recommends explicitly giving the model permission to say it doesn't know, which it says can drastically reduce false information (source). That's why this prompt asks for a "Check by ear" list instead of a guess:

<spellings>
[list every name, brand, product, and term in the video, spelled correctly,
for example: guest name, company name, show name, jargon]
</spellings>

<captions>
[paste the full .srt file here]
</captions>

Fix this caption file so it matches what was said.

Rules:
- Only fix words that were misheard or misspelled. Use the <spellings> list.
- Don't reword, shorten, or "improve" what anyone said. Don't fix their grammar.
- Don't change any caption numbers or timestamps.
- Keep the file in valid SubRip (.srt) format.
- If you aren't sure what was said in a line, leave it as-is and add it
  to a list called "Check by ear" with its caption number and timestamp.

Return the corrected .srt file first, then the "Check by ear" list,
then a list of every change you made (old text and new text).

Anthropic's prompting docs recommend putting long documents near the top of the prompt and wrapping material in XML-style tags so the model can tell it apart from your instructions (source), which is why the spellings and captions come first, in tags.

The list of changes matters. Read it. If the AI "fixed" something a person actually said, put it back.

4. Add the sounds that matter

Captions aren't only speech. The W3C says to transcribe relevant non-speech sound, and to include background music only if it's important to understanding the video. YouTube's own editor tip is the same idea: add text like [applause] or [thunder] so viewers know what's going on (source). The AI can't hear the video from a text file, so add these yourself while you check it in the next step.

5. Make the lines easy to read

The W3C's guidance for caption lines: one or two lines per caption, generally under 32 characters per line, a new sentence on a new line, and long sentences broken at a logical phrase (source). In the same chat, follow up with:

Now reflow the corrected captions for easier reading:
- One or two lines per caption, ideally under 32 characters per line.
- Start each new sentence on a new line.
- Break long sentences at a natural phrase, not in the middle of a name.
- Don't change any words, and keep every timestamp exactly as it is.

Then compare a few captions against the earlier version to make sure the words really didn't change.

6. Check it against the video yourself

This is the step you can't skip. Play the video with the corrected file and:

7. Upload it to YouTube

Per YouTube's help page, as of this post's date: sign in to YouTube Studio, select Subtitles from the left menu, select the video, click Add language and choose your language, then under "Subtitles" click Add. To use your file, select Upload file, choose With timing (your .srt has timestamps), choose the file, and select Save. You can also add captions during the upload process. Source: Add subtitles & captions (YouTube Help).

If you only have a clean transcript without timestamps, YouTube's Auto-sync option can time it to the video for you. YouTube says the transcript must be in the same language spoken in the video and a language its speech recognition supports, and it doesn't recommend transcripts for videos over an hour long or with poor audio (source).

To fix automatic captions directly instead, YouTube's steps are: in YouTube Studio, select Subtitles, click the video, click More next to the subtitles you want to edit, and edit or remove anything that wasn't transcribed properly (source).

Closed captions vs. captions burned into the video

The W3C explains that most captions are "closed," meaning viewers can turn them on or off, while "open" captions are always displayed and can't be turned off (source). Text burned into a short clip is the open kind. If you're cutting vertical clips, the same corrected file is your source text for on-screen words, so you fix the errors once and use the result everywhere. The clipping workflow is in How to Find the Best Short Clips in a Long Video With AI.

Mistakes to avoid

Official sources

Platform steps and rules in this guide are from the official pages above, accurate as of the published date. They can change, so check the linked pages before you rely on a specific step.