sxsphinxstack

Skills / Captions and subtitles

Captions and subtitles

Put accurate captions on the person's video. Auto-generate, fix every line by hand, style for phones, export or burn in. Use when they say "add captions", the auto-captions are wrong, or a video is going anywhere people watch muted.

Use this skill. Nothing to install.

Open in ChatGPT Open in Claude prefilled and ready; just hit send
for Gemini, Copilot, or Cursor: paste it, then say "use the captions and subtitles skill"

Don't have an agent?  ·  Raw file: skills/captions-and-subtitles.md

Watch the first two minutes

This is how a session goes when the auto-captions have already embarrassed you once. Click through it.

​(pastes the skill) youtube auto-captioned my video and my name is Ngozi but the captions say "and ghazi" the whole way through
Auto-captions are the draft, never the deliverable, and they reliably mangle exactly the words that matter most: names, project names, slang, technical terms. To the one audience that knows the topic, those errors read as carelessness. We fix every line by hand while the video plays.
every line? the video is six minutes
Every line, and it goes faster than it sounds because most lines need nothing. You're checking against the audio and stopping where it broke. Where does the video live right now: in an editor, on YouTube, or a bare file?
already uploaded to youtube
Then YouTube Studio, Subtitles: the auto track is editable in place. Fix the names, add sentence case and real punctuation since captions are reading material, and check that each caption appears as the words are spoken. Keep what was said: an "um" can go, but the meaning can't change.
honestly does this matter? people can just listen
Most short-form video is watched with sound off, and for deaf and hard-of-hearing viewers, captions are how they get the content at all. Wrong captions fail real people. When we're done, download the corrected SRT and keep it with the project; it travels with the video everywhere else it goes.

Scripted example of a real session.

Can't I just leave the auto-captions on?

Auto-captions get you a draft in minutes, and that is their whole job. They break on the words that carry your video: names, project names, technical terms, anything a dictionary hasn't met. Viewers who know the topic notice instantly, and viewers who depend on captions get a garbled version of your actual point. The skill keeps the fast part, auto-generating with whichever tool sits closest to where your video already lives, and adds the part that matters: a human pass over every line, with the video playing, until the captions say what you said.

What you end up with

Captions checked against the audio line by line, styled for phones, and shipped in the right format for each destination. Here is the caption pass from the session above.

Example
CAPTION PASS — SIX-MINUTE PROJECT DEMO
Draft: YouTube auto-captions  ·  Fixed in Studio, SRT downloaded after
FIXES 1
"and ghazi built this in react" → "Ngozi built this in React"
Name and proper noun casing
"the a p i key goes in dot e n v" → "the API key goes in .env"
Technical terms, the words the knowing audience reads first
"its kind of janky but it works" → kept word for word 2
Timing nudged 0.4 s earlier to land as it is spoken
One three-line wall split into two captions, about 40 characters per line
STYLE 3
Heavy sans-serif · white with black outline · two lines max · above the bottom quarter on the vertical cut · checked on a real phone at arm's length
SHIPPED
YouTube: corrected track published, SRT downloaded and saved with the project
Shorts cut: styled captions burned in, proofread before export because burn-in is final
  1. The fixes cluster where it hurts. Auto-captioning breaks on names and terms, which are exactly the words that mark you as careful or careless to people who know the topic.
  2. Captions transcribe. The phrasing stays yours, hesitations and all. A stray "um" can go; the meaning cannot change.
  3. The phone is the test device. Legibility gets verified by playing the export on an actual phone, at arm's length, since that is where it will be watched.

Questions people actually ask

What's the fastest way to get a draft?

Whichever tool sits closest to the video: CapCut and DaVinci Resolve auto-caption inside the editor, YouTube auto-captions after upload, and free Whisper-based tools transcribe a standalone file locally.

SRT file or burned-in captions?

Per destination. Sidecar SRT/VTT for YouTube and anywhere that accepts subtitle uploads, since viewers can toggle it and search can index it. Burn-in for TikTok, Reels, and Shorts, where styled captions are part of the video. Doing both costs little.

Can I clean up my grammar in the captions?

No. Captions transcribe what was said. An "um" can be dropped, but rephrasing yourself into someone more polished changes the record, and viewers lip-reading or half-listening notice the mismatch.

How big should captions be, and where on screen?

Two lines max, around 40 characters per line, big enough to read on a phone at arm's length. High contrast, white with a black outline or backing. Above the bottom quarter on vertical video, lower third on horizontal, never over faces or the thing being shown.

Are word-by-word pop animations okay?

Fine for shorts if your editor does them. Accuracy and legibility still rule: an animated caption that is wrong or unreadable is worse than a plain one that works.

Where to go from here

Every video skill here ends with a captions pass. The ones that send people to this page:

Edit a short Screen-recording tutorial Video resume

Captions done? Finish the upload properly:

Make a thumbnail Publish to YouTube Accessibility pass
Curious? Read the full skill — the exact instructions your agent gets 
---
name: captions-and-subtitles
category: media
description: Put accurate captions on the person's video. Auto-generate, fix every line by hand, style for phones, export or burn in. Use when they say "add captions", the auto-captions are wrong, or a video is going anywhere people watch muted.
---

# captions-and-subtitles

Caption a video with someone — theirs, of course. Most
short-form video is watched with sound off, and captions are how
deaf and hard-of-hearing viewers get the content at all, so wrong
captions fail real people, and so does shipping without them. The
session ends with captioned video exported or the corrected captions
live on the platform.

## Auto-generate, then fix by hand

Auto-captions are the draft, never the deliverable.

1. Generate: CapCut and DaVinci Resolve auto-caption inside the
   editor; YouTube auto-captions after upload (editable in Studio
   under Subtitles); for a standalone file, free Whisper-based
   tools transcribe locally. Use whichever sits closest to where
   the video already is.
2. Fix every line while the video plays. Auto-caption reliably
   mangles exactly the words that matter most: names, project
   names, slang, technical terms, anything not in a dictionary.
   Those errors read as carelessness to the one audience that
   knows the topic.
3. Punctuation and casing count — captions are reading material.
   Sentence case, real punctuation, no wall of caps.
4. Check timing: each caption appears as the words are spoken, not
   a beat late. Split any caption longer than two lines.
5. Keep what was said. Captions transcribe; they do not clean up
   grammar or swap in better phrasing. An "um" can go; the meaning
   cannot change.

## Styling

- Two lines max, roughly 40 characters per line.
- Big enough to read on a phone held at arm's length — test by
  playing the export on an actual phone, not by squinting at the
  desktop preview.
- High contrast: white text with a black outline or backing bar
  survives any footage. Skip fancy fonts; heavy sans-serif reads.
- Position in the safe zone: above the bottom quarter for vertical
  video (platform UI covers the bottom), lower third for
  horizontal. Never over faces or the thing being demonstrated.
- Word-by-word pop animations are fine for shorts if the editor
  does them; accuracy and legibility still rule.

## Sidecar file or burn-in

Two ways to ship captions; pick per destination:

- Sidecar (SRT/VTT): a separate subtitle file uploaded alongside
  the video. Viewers can toggle it, search indexes it, screen
  readers and translations can use it. This is the right choice
  for YouTube and anywhere that accepts subtitle uploads — export
  SRT from the editor, or download the corrected file from YouTube
  Studio.
- Burn-in: captions rendered into the pixels. Right for TikTok,
  Reels, and Shorts where styled captions are part of the video
  itself. Once burned, they cannot be turned off or fixed — final
  proofread happens before this export.

Doing both costs little: burn in for the short-form cut, keep the
SRT with the project for every other use.

## Done

- Every caption checked against the audio by a human — them
- Styled, positioned, and verified readable on a real phone
- SRT saved with the project, or burned-in export done, or the
  platform's captions corrected and published