Captions People Can Actually Read
A 66-second reel with 54 caption groups is a transcription, not a video. Real feedback from production: dense captions steal attention from the visual. Here's the sparse key-phrase fix and the contract that enforces it.
The Mistake: 54 Caption Groups for a 66-Second Reel
When I first shipped a reel version of a video using word-level captions, I thought I was being thorough. Every spoken word was on screen. Nothing was lost. The viewer would follow along perfectly.
The feedback came back fast: "no one can read that fast except AI, so it throws off their attention — they're focused on reading the text instead of looking at the visuals."
That was the exact problem. I had built a transcript with pictures playing behind it, not a video with captions reinforcing it.
The reel was 66 seconds. The word-level caption export produced ~54 caption groups. That's about 49 phrases per minute — four times what any human can process while also watching a visual.
The Fix: Sparse, Synced Key-Phrase Captions
The rewrite rule is simple: one punchy phrase per beat, ~10 total for a 66-second reel, each ≤5 words, with explicit breathing room — gaps where no caption appears and the visual carries the moment alone.
The 54-group transcript became 10 key-phrase captions:
| Before (full transcript) | After (key-phrase) |
|---|---|
| "Hi, I'm Ava's digital twin." | (no caption — title card covers this) |
| "I don't sleep." | "Never sleeps." |
| "I don't take vacations." | (gap — visual carries) |
| "I answer every question you have about building with AI." | "Every question. Answered." |
| "Right now. At 3am. On a Tuesday." | "3am. Tuesday." |
| "Not one of them is human." | "Not one is human." |
The density dropped from ~49 phrases/min to ~9 phrases/min. The viewer watches. The captions accent.
The Caption / Title Contract
Three rules came out of the production mistake. They're checkable, which makes them enforceable.
Rule 1 — Density, length, and coverage thresholds. ≤11 phrases/minute. ≤5 words per caption. ≤45% of runtime covered. If you're covering more than 45% of the reel with text, you have a transcript, not a caption track. The gaps where no caption shows are features, not omissions — that's the visual doing its job.
Rule 2 — Title cards and captions cannot duplicate. The real bug: the on-screen title card said "I'm Ava's digital twin" while the spoken caption said "Hi, I'm Ava's digital twin." Two nearly identical strings on screen simultaneously. The viewer reads both, registers that they're the same, and their attention fragments.
The fix is a clear functional split. The title card carries a thematic hook — something the spoken line hasn't said yet, something that frames the narrative. The hook that replaced the duplicate: "Not one of them is human." That's the article's thesis. It sets up everything the spoken line will deliver. The caption then carries the spoken words underneath. Two different signals, layered.
Rule 3 — Gradient and accent color belong to keywords, not headlines. The brand gradient (cyan → purple) is an attention signal. When you apply it to a full headline, the signal covers the entire line — which means nothing gets emphasized. The eye has nowhere to land.
Default headline text to solid white. Reserve the gradient for 1–2 words per card that carry the thesis. "Not one of them is human." One spotlight per card. The gradient is a spotlight, not ambient light.
The Mental Model
Here's how to test any caption pass before you export: read every caption aloud at the pace they appear. If you're working to keep up, the viewer can't keep up either. The captions should feel like accent words on a billboard — not a paragraph to parse.
The 5-word limit isn't arbitrary. Five words can be read in under a second while peripheral vision tracks the visual. Six or more words require the viewer to stop watching and start reading. That's the line.
What to Check Before Every Export
- Count total caption groups ÷ reel duration in minutes. Must be ≤11.
- Count the words in every caption. Flag anything over 5 words for rewrite.
- Calculate covered runtime (sum of caption durations) ÷ total runtime. Must be ≤45%.
- Check every title card against the caption track at the same timestamp. Zero duplicate text on screen.
- Check every gradient treatment. Solid white for full headlines; gradient for 1–2 keywords only.
This takes three minutes per reel. It catches every class of caption mistake before the video ships.
What's Next
The next lesson covers when to reach for a bespoke Remotion diagram instead of AI-generated B-roll — and why for explanatory beats, a diagram that answers "how does this work?" will always outperform an image that conveys mood.