How to Write Subtitles People Actually Read

A large share of video in a feed is watched with the sound off, which means the subtitles are not an accessibility extra sitting beside the content. For those viewers they are the content.

Most subtitles are bad in the same few ways, and all of them come from treating a transcript as if it were a subtitle file. They are different things.

A transcript is not a subtitle file

A transcript records what was said. A subtitle file decides what appears on screen, when it appears, and for how long, and those decisions are most of the work.

The rules that matter

Subtitling has a long professional tradition and most of it distils to a handful of practical limits. These are not arbitrary; each one exists because breaking it measurably slows reading.

• Two lines maximum on screen at once, never three
• Around 40 characters a line, fewer for vertical video where the frame is narrow
• At least one second on screen, even for a single word
• Break lines at natural phrases, never between an article and its noun
• One thought per subtitle rather than one sentence, if the sentence is long

The line break rule is the one that separates subtitles that feel effortless from ones that feel like work. Breaking after "the" or "a" forces the reader to hold an incomplete phrase across two lines, and it is noticeable even when nobody can say why.

Timing

A subtitle should appear as the words are spoken and stay long enough to be read comfortably. Reading is slower than listening, so a subtitle that vanishes the instant the speaker stops is too fast.

1. Start it on the first word rather than slightly ahead, which spoils jokes and reveals
2. Hold it a beat past the last word so the end of the line can be read
3. Never let one subtitle span a scene change, which breaks the association
4. Merge very short consecutive lines rather than flashing two half second cards

Automatic captions as a starting point

Generated captions are a reasonable first draft and a poor final one. They get ordinary speech right most of the time and fail predictably on the things that matter most: names, technical terms, numbers, and anywhere two people talk over each other.

The efficient approach is to pull the automatic transcript, then edit it rather than typing from nothing. The transcript tool on this site retrieves the existing captions from a video, preferring a human written track over an automatic one where both exist, and gives you plain text or a timed SRT.

Then fix, in this order: proper nouns first, because a wrong name is the error readers notice; then numbers; then punctuation, which the automatic version mostly guesses; then line breaks.

Burned in or separate?

Burned in subtitles are part of the picture. They always appear, they cannot be switched off, they cannot be translated, and no search engine can read them.

A separate subtitle file can be toggled, translated, styled by the player, and indexed. For short form where autoplay is silent and the viewer will not tap a button, burned in usually wins. For anything longer, a real subtitle file is worth more, and there is nothing stopping a video having both.