| name | narration-slide-division |
| description | Split content between the voice and the slide so they complement instead of repeat. Use at node 03 (script / cue-plan) AND node 04 (visual / slide copy) when authoring any beat that has both narration and on-screen text — the anti-duplication contract that keeps the voice from reading the slide aloud. |
Narration ↔ Slide Division
The content-side sibling of render-lab-design-system's "Quality bar": that skill answers "is this good to look at?"; this one answers "is the voice doing a different job from the slide?". It is one contract read by BOTH node 03 (narration / cue plan) and node 04 (slide copy) — the division is a relationship between two artifacts, so neither single-artifact skill can own it. Caption emphasis text stays governed by caption-and-emphasis (it already forbids captions that repeat narration); this skill governs the slide text / title / slot strings vs the cue narration.
The spine
The slide carries the NOUN. The voice carries the VERB / the WHY.
- SLIDE = the nouns to SEE and remember — product name, number, model list, price, label, the one turn-word, the real UI. Things that are wrong to merely hear and right to read. Concrete, scannable, permanent.
- VOICE = the connective tissue to FEEL/UNDERSTAND but not read — motivation, the so-what, the causal bridge between beats, the stance/judgment, the segue into the next idea. Transient, conversational, forward-moving.
Anti-duplication gate (the one to enforce)
This is the Redundancy Principle (Mayer & Clark, e-Learning and the Science of Instruction, ch.7): identical on-screen text + narration lowers comprehension — the eye can process one visual at a time, and the viewer wastes effort cross-checking text against voice.
The delete-the-overlap test, per beat:
Delete every content word present in BOTH the slide string and the narration string for the same beat. If the narration still stands and adds something the slide doesn't → pass. If it collapses to nothing, the voice was just reading the slide → rewrite it (paraphrase/extend it, or point to it and add the why — never recite it).
Run the same test against the caption track: caption ≈ narration → drop the caption (caption-and-emphasis).
Allowed overlap (the only exceptions)
- Brand/product name + the final CTA tag — say it while it's on screen; naming reinforces. Once at reveal and once at the endcard, not every beat.
- A short keyword memory-hook shorter than the narration (number "300+" on screen while the voice says "and counting"). A label, never the full sentence.
- Technical/unfamiliar terms the ear can't catch — a model id, a price string, an API shape. Show them.
- A pure kinetic-type / no-graphic beat where text IS the content (voice off or sparse) — nothing else to look at, so no redundancy.
Narration craft — how to talk
- Conversational, spoken-not-written. Contractions, address "you", read it aloud before locking. Cut "essay voice."
- Sentences under ~15 words; vary length so it isn't monotone. Ban filler: "just / really / basically / in order to / due to the fact that."
- Pace ~130–150 wpm (numbers/prices/URLs read slower and eat extra budget).
- Word budget per beat ≈ 2.5 spoken words per second of beat duration, then subtract for reveal silence. A 4s beat ≈ 8–10 words, not 25. Leave deliberate silence on a big reveal so the visual lands.
- Add what the slide can't show — the motivation/stakes ("five things that can break at three in the morning" over five logos), the causal bridge ("so we collapsed all of it into one"), the so-what of a number (the judgment, not the digits).
- Open on tension, end fast. Hook in the first 1–3s; one idea per beat; CTA is one specific action, not a summary.
- One thing at a time, and the visual lands ON the spoken word — a ≥0.5s lag reads amateur.
Slide copy — what to write
- One headline + optional one support line per beat. One idea per screen. Headline noticeably larger/heavier; one weight contrast, never Bold+Italic+Caps stacked.
- Headline is a NOUN PHRASE or a single turn-line, not a sentence the VO will speak. "300+ models" beats "We support over three hundred models."
- Rule of six (ceiling, not target): ≤6 words/line, ≤6 lines/screen — aim well under; for kinetic promo, one phrase of 3–7 words per beat.
- A number or label beats a sentence when the value IS the point — prices, counts, deltas, model ids show as the artifact; the voice says why.
- Readability dwell: hold text still ≥1s per 13 characters (≈2.3s for a 30-char line). Animate IN fast, then HOLD for the read — never animate during the read.
- Real UI copy only — never invent labels. Mirror the actual product screen. Abstract the periphery; make the core workflow the only legible element.
How to apply in this pipeline
Per beat, author and carry forward three short lines (node 04 declares them; shot-brief-writer records them; QC checks them):
slide carries: <nouns / number / turn-word>
voice carries: <motivation / mechanism / bridge — the why>
overlap allowed: <brand | cta | keyword:"300+" | none>
Then run the delete-the-overlap test before locking the beat. If voice carries: is just a restatement of slide carries:, the beat fails — rewrite the voice to the motivation, mechanism, or bridge.
