AI Avatar Scripts & Prompts: HeyGen and Synthesia (2026)
AI avatars stopped being a novelty between last year’s demos and this year’s Q4 calendar. In 2026, a talking presenter generated on HeyGen or Synthesia is a normal production asset: Black Friday 11⁄27 offers, Christmas 12⁄25 greetings, New Year 1⁄1 recaps, onboarding explainers, localized product drops. The avatar tech is solved; what isn’t solved is the part that persuades: the script.
Most avatar videos fail in the first three seconds because someone pasted in copy written for the eye — long clauses, stacked adjectives, a paragraph that reads fine on a landing page and dies out loud. The second failure is visual: a static talking head with no B-roll, no framing, no rhythm. An avatar video is still a video. It needs shot grammar.
This guide treats the avatar clip as what it is — spoken performance plus a visual track — with script structure, direction prompts and B-roll templates.
The script is the prompt
Think of the avatar stack as two prompt layers:
- The spoken layer — the script: the input for pacing, emphasis and delivery. Every word is a direction to the voice.
- The visual layer — avatar, framing, gesture tone, B-roll, captions. This is where shot grammar lives.
HeyGen and Synthesia handle mechanics — avatar, voice, captions, multi-language versions — but neither fixes a flat script. Treat the script file like a video prompt: structured, timed, written for one outcome.
Write for the ear, not the page
Spoken copy follows different rules:
- Hook in the first sentence: the viewer decides before the intro ends. Open with the problem or claim, not a greeting.
- One idea per sentence: if it needs a comma stack, split it. 12–18 words per sentence is a good spoken range.
- Mark the beats:
[pause],[slow],[emphasis]written into the script file. Even if a tool ignores them, they make your read testable. - Numbers spelled out in rhythm: “forty percent off — this week only” lands; “40% off (limited time)” doesn’t survive synthesis.
- End on the ask: one action per script — shop the offer, book the demo. Not two.
A workable spine for short promos: hook (3s) → problem (5s) → product as answer (10s) → proof/feature (8s) → CTA (4s). Roughly 30 seconds, 75–85 words. Adjust the spine, not the discipline.
The seven shot elements for avatar video
Talking-head footage still obeys the framework — the weight shifts toward sound and timing:
| Element | What to specify | Example |
|---|---|---|
| Subject | Avatar archetype + wardrobe + energy level | “presenter in a charcoal knit sweater, warm but brisk energy” |
| Framing | Head-and-shoulders crop, eye line, headroom | “medium close-up, eyes upper third, even headroom” |
| Camera | Static or subtle movement; distraction is the enemy | “locked-off frame” |
| Motion | Gesture language and facial emphasis | “small gestures on key points, one eyebrow lift on the hook” |
| Light & color | Flat, flattering, brand-neutral lighting | “soft front-left key, neutral background” |
| Timing | Words-per-minute feel and beat markers | “brisk 150 wpm, [pause] before the offer, [slow] on the price” |
| Sound | The whole show: voice, pacing, music bed | “confident mid voice, music bed at -20 dB, burned-in captions” |
The sound row is no longer a post note — the voice is the generation. Write voice direction into the script header.
Copy-paste templates
1. 30-second product promo script + direction (9:16, TikTok/Reels/Shorts)
FORMAT: 9:16 vertical, 30 seconds, avatar + B-roll overlay, burned-in captions.
AVATAR DIRECTION: presenter in a plain sage-green tee, warm but brisk energy, medium close-up, locked-off
frame, soft front-left key light, clean neutral background. Small hand gestures on key points.
VOICE: confident mid-range voice, brisk pace (~150 wpm), [pause] before the offer.
SCRIPT: Your coffee is cold again. [pause] It doesn't have to be. This insulated travel mug keeps every
sip hot for hours — double-wall steel, leak-proof lid, fits your car cup holder. [slow] Today it's
thirty percent off. Tap the link and grab yours before the price goes back up.
B-ROLL PLAN: 0-3s close-up of steam rising from a mug; 6-12s macro of the lid sealing shut; 18-24s the
mug in a car cup holder at golden hour. CTA card from 27-30s with product and price.
2. 60-second explainer script + direction (16:9, YouTube / website)
FORMAT: 16:9 horizontal, 60 seconds, avatar A-roll with two B-roll cutaways, captions on.
AVATAR DIRECTION: presenter in a navy oxford shirt, calm teacher energy, medium shot, subtle 5% push-in
over the full minute, soft key + gentle background gradient.
VOICE: warm, unhurried (~140 wpm), [emphasis] on the three steps, [pause] before the CTA.
SCRIPT: Most teams lose hours every week re-cutting the same video for every platform. Here's the faster
way. [emphasis] Step one: pick one hero video you already made. [emphasis] Step two: extract its camera
and timing structure — what the shot does, not what it says. [emphasis] Step three: rebuild that
structure once, then re-skin it for each channel. Three steps, one workflow, and your next campaign ships
in an afternoon. [pause] See how it works at the link below.
B-ROLL PLAN: 12-20s screen capture of a timeline being re-cut (generic UI, no real logos); 34-44s
split-screen of one clip re-skinned into two aspect ratios (9:16 + 16:9). End card 55-60s.
3. Holiday greeting + B-roll prompts (9:16, Christmas → New Year)
FORMAT: 9:16 vertical, 20 seconds, avatar greeting + generated B-roll, warm color grade, captions on.
AVATAR DIRECTION: presenter in a dark red knit sweater, genuinely warm energy, medium close-up, slow
10% push-in, festive but tasteful background (plain warm wall, soft bokeh string lights, no branded decor).
VOICE: warm, slower pace (~125 wpm), smile audible on the final line.
SCRIPT: From all of us to all of you — thank you for being part of our year. [pause] We're taking a short
break, then we're back on January second with more of what you asked for. [slow] Happy holidays, and
see you in the new year.
B-ROLL PROMPTS (generate separately, cut over the script):
- "Close-up of string lights glowing on a dark wall, warm bokeh, slow rack focus, 3 seconds, 9:16."
- "Snowfall drifting past a window at night with warm interior light reflecting on the glass, gentle
motion, 4 seconds, 9:16."
- "Golden confetti falling slowly against a dark warm background, shallow depth of field, 3 seconds, 9:16."
Before you generate: checklist
- [ ] Hook is in the first sentence — no greeting, no throat-clearing
- [ ] Sentences 12–18 words, one idea each; read the script aloud before generating
- [ ] Beat markers written in: [pause], [slow], [emphasis]
- [ ] Script length matches the format: ~75–85 words for 30s, ~140–150 for 60s
- [ ] Avatar direction stated: wardrobe, energy, framing, lighting
- [ ] B-roll planned with durations and aspect ratio (9:16 / 16:9)
- [ ] One CTA only, at the end
- [ ] No brand claims, borrowed scripts or licensed music in the package
Common mistakes and FAQ
“The avatar sounds bored.” The script is flat, not the voice: cut adjective stacks, add beat markers, give the line a reason to be said now — urgency reads as energy.
“Do I write the B-roll prompts differently from normal video prompts?” Same seven elements, shorter durations: 2–5 second shots that don’t compete with the spoken line, matched to the A-roll’s color temperature so cuts don’t jar.
“Can I localize the video for other languages?” Both platforms generate multi-language versions, but direct translation sounds written. Re-time each language’s rhythm and re-check the hook — wordplay rarely survives.
“How much does the avatar’s appearance matter?” More than people expect, less than the script: wardrobe and background must not fight your brand palette; beyond that, credibility rides on voice clarity and script relevance.
“Is it fine to imitate a competitor’s video structure?” Structure is grammar — problem-hook, proof, CTA is everyone’s. Their script, likeness, music and claims are theirs. Lift the skeleton from any reference, write every word yourself, keep extraction notes.
An avatar is a delivery device; the script is the product. Write for the ear, direct the frame, and your next 30 seconds gets watched — pull a reference clip apart with SocialToPrompt and rebuild its hook, beats and B-roll as your own script.
Share Article