Captions are not decoration. On a phone, in a noisy room, with the sound off by default on most feeds, they are the primary channel. A caption that arrives too fast, wraps badly, or sits at 30 percent contrast is a caption nobody finishes — and an unfinished caption is an unfinished video.

The good news is that readability is measurable. Four numbers cover most of it, and none of them require taste to apply.

Start with characters per second

Reading speed for on-screen text is typically estimated at 15 to 20 characters per second for comfortable comprehension, with subtitling standards converging somewhere in that range. Below about 12 characters per second the text feels sluggish and out of sync with speech. Above about 20, viewers start skipping words, and once they skip they tend to stop reading entirely.

For short-form video the practical target is 15 to 17 characters per second. It is slightly conservative on purpose: a viewer scrolling in a moving car, or reading in a second language, needs the margin.

The test is simple. Find the longest caption on screen, count its characters, and divide by how many seconds it is visible. If the answer is over 20, either the line is too long or it is on screen for too short a time.

Line breaks are editorial decisions

Automatic line wrapping breaks wherever the box runs out of room, which is usually mid-phrase. That small failure costs real comprehension, because the eye treats each line as a unit and a line that ends in the middle of a clause forces a re-read.

Break captions at clause boundaries instead:

Before a conjunction: "we tried that / but the audio was unusable".
After a complete subject-verb-object unit.
Before a prepositional phrase that qualifies the whole clause.
Never between an article and its noun, or between a number and its unit.

Two lines, and prefer a shorter first line. A top-heavy caption reads as a block; a bottom-heavy one reads as a sentence continuing.

Two lines, 32 to 42 characters

Three-line captions are the most common readability failure in short-form video, and they usually appear for a structural reason: the speaker said a long sentence and nobody split it.

Split it. A caption that occupies three lines at a large size either covers the subject's face or forces a smaller font size, and small text is worse than extra screens. The usable range is roughly 32 to 42 characters per line at a size that is legible on a 6-inch screen — which is really a statement about height, not width: a caption should be at least about 4 percent of the frame height per line.

Word-by-word or sentence-level?

Both work, for different jobs.

Word-by-word — one or two words at a time, with the active word highlighted — is energetic and pairs well with fast, punchy delivery. It holds attention because the frame keeps changing. It struggles with long or complex sentences, because a viewer cannot look ahead.

Sentence-level — a full clause at a time — is calmer and better for explanation, story and anything with numbers, names or jargon.

A useful compromise, and the one most creators settle on, is short phrases of three to five words with the active word highlighted. It keeps the motion of word-by-word captions and the context of a full sentence.

Contrast decides whether any of it is legible

A dark caption on a mid-tone background is invisible regardless of length. Three approaches, in order of robustness:

A background plate. A semi-transparent box behind the text. Ugly on some designs, unbeatable for readability, and the right choice for b-roll-heavy footage.
Outline plus shadow. A solid stroke of roughly 4 to 8 percent of the font size, plus a soft shadow. This is what most broadcast captions use, and it survives almost any background.
A solid colour with no treatment. Only safe when you control the background, which in short-form video you usually do not.

Check contrast against the brightest area of the frame behind the text, not the average. A white caption over a white shirt is a caption that disappears for two seconds.

Caption the words, not every sound

There is a temptation to caption every "um", every laugh, every "(music)". For accessibility subtitles this is correct practice. For short-form video it is usually noise, and it inflates the character count you are trying to keep down.

The sensible line: caption all dialogue, drop filler sounds, and keep a sound description only when the sound is the point — a door slamming, a phone ringing, a room going silent.

Fix the transcript where it matters

Automatic transcription fails predictably. Check these categories in this order:

Names of people, companies and products. Most likely to be wrong, most likely to be noticed.
Numbers, prices and dates. A wrong digit is a factual error, not a typo.
Jargon and acronyms. Often transcribed phonetically into nonsense.
Homophones in context. "Their" and "there", "principle" and "principal".

Everything else — a dropped filler word, one casing error — rarely changes meaning and is not worth the time before publishing.

Reading speed is a property of the whole clip

A viewer who has to read ten captions in fifteen seconds is not watching a video any more; they are reading a document at speed. When a clip feels frantic even though the editing is calm, the cause is usually caption density, not cuts.

Two levers: remove words, or add time. Removing words means cutting a repeated phrase or a hedge — "I think that, basically, what we found was" becomes "what we found was". Adding time means leaving a beat of silence before the next caption, which also makes the clip feel less breathless.

Accessibility is not a separate task

Captions written for readability are close to captions written for accessibility, with two additions worth making:

Speaker labels when two people are on screen and the voices are similar.
Accurate timing, so a viewer who is deaf can follow who said what, not just the words.

Everything else — two lines, sensible breaks, adequate contrast, corrected names — is shared between the two goals. There is no trade-off to manage here; the readable caption is the accessible one.

A pass/fail check before export

Longest caption is at or below 20 characters per second; most are 15 to 17.
No caption exceeds two lines.
No line break splits a clause, an article from its noun, or a number from its unit.
The caption clears the platform interface: inside the middle 70 percent of the width, above the bottom 20 percent.
Contrast is legible over the brightest background frame in the clip.
Names, numbers and jargon are correct.
Reading the captions alone tells you what the video says.

If all seven pass, the captions will do their job in a feed where nobody can hear you.

How to Write Captions People Actually Read