How to prompt HeyGen

Prompting HeyGen is scriptwriting plus casting. On the avatar models the prompt is not a scene description — it is the exact words the presenter will speak, read by the voice you pick or by audio you attach. Delivery is directed through the writing itself: sentence length sets pacing, punctuation sets pauses, spelling sets pronunciation. There is no duration setting — the clip lasts as long as the script does. Everything visual (who appears, framing, resolution) is a setting, not prompt text.

Anatomy of a HeyGen Avatars prompt

1
The script is the prompt
Every character you type is spoken verbatim. Write for the ear, not the eye: contractions, short words, the rhythm of talk. If a line feels stiff read aloud, it will sound stiff on the avatar.
2
Hook sentence
Open with one short, direct line — a question, a claim, a number. The first two seconds decide whether a feed viewer stays; do not spend them on "Hi, welcome to this video".
3
One idea per sentence
Full stops are where the avatar breathes and resets its inflection. Many short sentences read more naturally than a few long ones; commas give shorter breaks, periods give longer ones with downward inflection.
4
Pauses via punctuation
There is no pause tag in the script — punctuation is the pause system. An ellipsis ("...") adds a beat of air before a payoff; a new sentence adds a full stop's pause.
5
Pronunciation spelling
Hard names and acronyms are fixed in the spelling: "loo-MAY-ah" for Lumeya, "A-P-I" for API. Respell every occurrence — the engine reads what is written, each time.
6
Casting (settings, not text)
Who appears and how it sounds are parameters: avatar name, voice name, or an attached audio track that overrides the built-in TTS entirely. Keep them out of the script text.

Template

[hook — one short spoken sentence]. [point one, a single sentence]. [point two]. [proof or number, written out as words]... [payoff after the beat]. [CTA — say exactly what to do next].

10 example prompts that work

Product explainer, 25-30 seconds

16:9 · Avatar V5 · 1080p · voice: Warm Pro Narrator

Ever spent a whole weekend editing one video? There is a faster way. Cliply turns raw clips into finished ads. You upload the footage. It writes, cuts, and captions. Done in minutes... not days. Try it free this week. The link is below.

About 45 words lands near 20 seconds at a natural pace. Every sentence carries one idea, so the avatar gets clean pauses; the ellipsis before "not days" gives the contrast a beat to land.

Vertical UGC-style ad from a face photo

9:16 · Avatar 4 (face image) · talking style: expressive · 720p

Okay, real talk. I have tried every planner app out there. This one actually stuck. It takes thirty seconds every morning... and my whole day just works. I am not going back.

Face-image mode animates any portrait you upload — your founder, your customer, a generated character. The "expressive" talking style loosens the delivery for UGC energy; "stable" is the calmer default.

Brand name pronunciation fix

9:16 · Avatar V5 · 1080p

Say hello to loo-MAY-ah. loo-MAY-ah reads your sleep data and tells you one thing to change. One thing. Every morning.

Phonetic respelling with hyphens is the reliable pronunciation fix — and it must appear at every occurrence, because the script is read literally each time. Trade-off: if you enable SRT captions, the respelling shows up in them too.

Acronyms and numbers written for the ear

16:9 · Avatar 4 · 1080p

Our A-P-I handles ten thousand requests a second. Setup takes five minutes. The first one thousand calls are free.

Write acronyms hyphenated ("A-P-I") and numbers as words — "10,000" can be read many ways, "ten thousand" only one. This is the single cheapest quality upgrade for spoken scripts.

Transparent presenter overlay for a course

16:9 · Avatar V5 · transparent background on · format: webm · fit: contain

In this lesson we will set up your first automation. Watch the left side of the screen. I will walk you through every click.

Avatar V5 can render with a transparent background when you export WebM — drop the file over a screen recording in any editor and you have a corner presenter without a green screen. Use fit "contain" so the framing is not cropped.

Your own voiceover, avatar lip-syncs

9:16 · Avatar V5 · audio_url attached · 1080p

Workflow: generate the voiceover first — for example with ElevenLabs on Clipwave — then paste its URL into the Audio URL field and pick your avatar. The attached audio replaces the built-in TTS and the lips follow your track exactly.

This is the control path: when the vocal performance matters more than convenience (a cloned voice, a directed read, exact pacing), lock the audio first and let the avatar act to it.

One-prompt brief for the Video Agent

9:16 · V3 Video Agent · avatar: auto · voice: auto

Make a 45-second vertical ad for a home espresso machine called Brewdeck, aimed at remote workers. Tone: dry, confident, lightly funny. Cover: heats in forty seconds, fits under kitchen cabinets, one-button cleanup. End on the line "Your commute is now four steps." Use the attached product photos.

The Video Agent takes a brief, not a verbatim script — it writes the script and, with avatar and voice set to "auto", casts the presenter itself. The V3 agent accepts up to 20 attached reference files.

Re-voice an existing video (Lipsync)

Lipsync Precision · match audio length: on

Input: a finished 30-second talking-head ad with an outdated offer. Attach the new voiceover with the corrected lines as the audio URL. Turn on "Match audio length" so the clip adapts to the new track, and "Strip source music" if the original had a bed you plan to replace.

Lipsync replaces the audio on footage you already have and re-syncs the mouth to it — no re-shoot. Precision re-infers the speaker for the tightest sync; Speed is the lighter, faster pass for drafts.

Localize a winning ad (Translate)

Translate Precision · output language: es / de / pt

Input: your best-performing English ad. Run Translate with output language Spanish, then again for German and Portuguese. The voice is re-generated in the target language and the lips are re-synced to it.

Eleven output languages are available (English, Spanish, French, German, Italian, Portuguese, Japanese, Chinese, Hindi, Arabic, Korean). Use Speed for internal review and Precision for the version you ship. Burned-in on-screen text is not translated — keep text out of the master.

Feed-format explainer with captions

4:5 · Avatar 4 · 1080p · SRT captions: on

Hiring is broken. Forty percent of new hires quit inside a year. We built Scout to fix the first ninety days... not the interview. Scout writes the onboarding plan before day one. Your new hire opens one link — and knows exactly what winning looks like. Try Scout on your next hire.

4:5 fills more of the feed than 16:9; the SRT caption toggle covers sound-off viewers. Numbers are written as words ("Forty percent") so the read is unambiguous and the captions match the speech.

Settings that matter

Avatar
Pick one presenter and keep it across a series — the recurring face becomes the brand asset. "Abigail Sofa Front" is the default; the Video Agents also accept "auto" and cast for you.
Voice
Match the voice to market and format ("Warm Pro Narrator" is the default). When no built-in voice fits, attach your own track as Audio URL — attached audio always overrides the script TTS.
Resolution
720p for drafts, 1080p to ship. 4K exists only on Avatar V5 — reserve it for hero placements and big screens.
Aspect ratio
16:9 for YouTube and web, 9:16 for Reels/TikTok, 4:5 or 5:4 for feeds, 1:1 for tiles. The Video Agents are landscape-or-portrait only.
Captions
Toggle SRT captions for sound-off placements. Remember phonetic respellings in the script will appear in the caption text too.
Output format
MP4 by default; switch to WebM on Avatar V5 when you enable the transparent background — WebM is what carries the alpha channel.

Do and don't

Do

Read the script aloud before generating — if you run out of breath mid-sentence, the avatar will too.
End every idea with a full stop; that is where the pause and the inflection reset happen.
Write numbers, prices and acronyms the way they should be spoken ("ten thousand", "A-P-I").
Respell hard names phonetically with hyphens, at every occurrence ("loo-MAY-ah").
Plan length by word count: roughly 130-160 spoken words per minute.
Generate the voiceover first and attach it when the vocal performance matters more than convenience.

Don't

Don't paste written prose — text written for the eye sounds stiff out loud; rewrite it in spoken register.
Don't put stage directions in the script ("[smiles]", "*points at camera*") — every character you type is spoken aloud.
Don't look for a duration setting — length is the script length (or the attached audio); cut words, not settings.
Don't stack nested clauses and parentheticals — they trip the pacing; split into two sentences.
Don't fix a mispronunciation by regenerating and hoping — respelling fixes it; retries rarely do.

Advanced techniques

Casting a recurring presenter

A spokesperson works because it repeats. Choose one avatar and one voice, note them down, and reuse the pair on every video in the series — audiences read the consistency as a real brand presenter. Digital-twin avatars (V3/4/5) are pre-built, studio-quality presenters; face-image mode instead animates any portrait you upload, which is how you turn a founder photo or a generated character into the face of the channel.

Audio-first workflows

There are three ways to drive the mouth, in increasing order of control: type a script and let the built-in TTS voice read it (fastest); attach your own voiceover as Audio URL, which overrides the TTS completely and makes the avatar act to your exact performance; or run Lipsync on footage that already exists to replace its audio and re-sync the lips — the path for correcting an offer, updating a claim, or re-voicing a real human video without a re-shoot.

The localization pipeline

Master the ad once, in your strongest language, then run Translate per market instead of rewriting and regenerating from scratch. The dub re-generates the voice in the target language and re-syncs the lips, so the performance survives translation. Two practical rules: keep burned-in text out of the master, because Translate only touches audio and lips — on-screen text stays in the source language; and use the Speed tier to review a market before committing the Precision pass to the version you publish.

Writing a brief for the Video Agent

The Video Agent inverts the contract: you hand it a brief and it writes the script, picks the shots and — with avatar and voice on "auto" — casts the presenter. A brief that works names five things: what the product is, who the video is for, the tone in two or three adjectives, the two or three points that must be covered, and any mandatory line verbatim. Attach reference files (the V3 agent takes up to 20) so claims and visuals come from your material. The V2 agent runs at fixed 30, 60 or 90-second lengths; V3 paces itself.

Which variant to use

HeyGen Avatar V5
Default digital twin. Up to 4K, transparent background, MP4 or WebM output — the full feature set.
HeyGen Avatar 4
Solid 1080p digital twin when you need none of the V5 extras.
HeyGen Avatar 3 (budget)
Fast, low-cost drafts of the same script before committing the final render to V4/V5.
HeyGen Avatar 4 (face image)
Animate any portrait photo you upload; choose stable or expressive delivery.
HeyGen V3 Video Agent
One brief, no script: the agent writes, casts avatar + voice, and can use up to 20 reference files.
HeyGen V2 Video Agent
Brief-to-video at fixed 30/60/90-second lengths when you need a predictable duration.
HeyGen Lipsync Precision / Speed
Replace the audio on an existing video and re-sync the lips — Precision to ship, Speed to draft.
HeyGen Translate Precision / Speed
Dub a finished video into one of 11 languages with re-synced lips.

Frequently asked

How do I control the video length?

You don't set it — you write it. The clip lasts exactly as long as the spoken script (or the attached audio). Budget 130-160 words per minute: a 75-word script lands near 30 seconds. The one exception is the V2 Video Agent, which offers fixed 30/60/90-second lengths.

Can the avatar speak with my own voice?

Yes. Attach any audio track as the Audio URL — your recorded voice, a cloned voice, a directed studio read — and it overrides the built-in TTS completely. The avatar lip-syncs to your track.

How do I fix a mispronounced word?

Respell it phonetically in the script, with hyphens, everywhere it occurs: "loo-MAY-ah", "A-P-I", "SEE-kwel". The engine reads exactly what is written. If SRT captions are on, the respelling appears in them — accept it or edit the caption file after.

Can I put the presenter on top of my own footage?

Yes, on Avatar V5: enable the transparent background and export WebM, which carries the alpha channel. Overlay the file on a screen recording or b-roll in any editor — an instant corner presenter for courses and demos.

What is the difference between the Avatar models and the Video Agent?

Avatar models speak your script verbatim — you control every word. The Video Agent takes a brief, writes its own script, and with "auto" settings casts the avatar and voice too. Use avatars when the copy is locked, the agent when you want a finished draft from one paragraph.

How do I dub a video into another language?

Run it through Translate with the target output language — 11 are available, from Spanish to Japanese. The voice is re-generated in that language and the lips re-synced. Burned-in on-screen text is not translated, so keep text overlays out of the master you localize.

Try HeyGen Avatars in Media Forge →About HeyGen Avatars All models

How to prompt HeyGen

Anatomy of a HeyGen Avatars prompt

1
The script is the prompt
Every character you type is spoken verbatim. Write for the ear, not the eye: contractions, short words, the rhythm of talk. If a line feels stiff read aloud, it will sound stiff on the avatar.
2
Hook sentence
Open with one short, direct line — a question, a claim, a number. The first two seconds decide whether a feed viewer stays; do not spend them on "Hi, welcome to this video".
3
One idea per sentence
Full stops are where the avatar breathes and resets its inflection. Many short sentences read more naturally than a few long ones; commas give shorter breaks, periods give longer ones with downward inflection.
4
Pauses via punctuation
There is no pause tag in the script — punctuation is the pause system. An ellipsis ("...") adds a beat of air before a payoff; a new sentence adds a full stop's pause.
5
Pronunciation spelling
Hard names and acronyms are fixed in the spelling: "loo-MAY-ah" for Lumeya, "A-P-I" for API. Respell every occurrence — the engine reads what is written, each time.
6
Casting (settings, not text)
Who appears and how it sounds are parameters: avatar name, voice name, or an attached audio track that overrides the built-in TTS entirely. Keep them out of the script text.

Template

[hook — one short spoken sentence]. [point one, a single sentence]. [point two]. [proof or number, written out as words]... [payoff after the beat]. [CTA — say exactly what to do next].

10 example prompts that work

Product explainer, 25-30 seconds

16:9 · Avatar V5 · 1080p · voice: Warm Pro Narrator

Ever spent a whole weekend editing one video? There is a faster way. Cliply turns raw clips into finished ads. You upload the footage. It writes, cuts, and captions. Done in minutes... not days. Try it free this week. The link is below.

About 45 words lands near 20 seconds at a natural pace. Every sentence carries one idea, so the avatar gets clean pauses; the ellipsis before "not days" gives the contrast a beat to land.

Vertical UGC-style ad from a face photo

9:16 · Avatar 4 (face image) · talking style: expressive · 720p

Okay, real talk. I have tried every planner app out there. This one actually stuck. It takes thirty seconds every morning... and my whole day just works. I am not going back.

Brand name pronunciation fix

9:16 · Avatar V5 · 1080p

Say hello to loo-MAY-ah. loo-MAY-ah reads your sleep data and tells you one thing to change. One thing. Every morning.

Acronyms and numbers written for the ear

16:9 · Avatar 4 · 1080p

Our A-P-I handles ten thousand requests a second. Setup takes five minutes. The first one thousand calls are free.

Write acronyms hyphenated ("A-P-I") and numbers as words — "10,000" can be read many ways, "ten thousand" only one. This is the single cheapest quality upgrade for spoken scripts.

Transparent presenter overlay for a course

16:9 · Avatar V5 · transparent background on · format: webm · fit: contain

In this lesson we will set up your first automation. Watch the left side of the screen. I will walk you through every click.

Your own voiceover, avatar lip-syncs

9:16 · Avatar V5 · audio_url attached · 1080p

Workflow: generate the voiceover first — for example with ElevenLabs on Clipwave — then paste its URL into the Audio URL field and pick your avatar. The attached audio replaces the built-in TTS and the lips follow your track exactly.

This is the control path: when the vocal performance matters more than convenience (a cloned voice, a directed read, exact pacing), lock the audio first and let the avatar act to it.

One-prompt brief for the Video Agent

9:16 · V3 Video Agent · avatar: auto · voice: auto

Make a 45-second vertical ad for a home espresso machine called Brewdeck, aimed at remote workers. Tone: dry, confident, lightly funny. Cover: heats in forty seconds, fits under kitchen cabinets, one-button cleanup. End on the line "Your commute is now four steps." Use the attached product photos.

Re-voice an existing video (Lipsync)

Lipsync Precision · match audio length: on

Input: a finished 30-second talking-head ad with an outdated offer. Attach the new voiceover with the corrected lines as the audio URL. Turn on "Match audio length" so the clip adapts to the new track, and "Strip source music" if the original had a bed you plan to replace.

Localize a winning ad (Translate)

Translate Precision · output language: es / de / pt

Input: your best-performing English ad. Run Translate with output language Spanish, then again for German and Portuguese. The voice is re-generated in the target language and the lips are re-synced to it.

Feed-format explainer with captions

4:5 · Avatar 4 · 1080p · SRT captions: on

Hiring is broken. Forty percent of new hires quit inside a year. We built Scout to fix the first ninety days... not the interview. Scout writes the onboarding plan before day one. Your new hire opens one link — and knows exactly what winning looks like. Try Scout on your next hire.

4:5 fills more of the feed than 16:9; the SRT caption toggle covers sound-off viewers. Numbers are written as words ("Forty percent") so the read is unambiguous and the captions match the speech.

Settings that matter

Avatar
Pick one presenter and keep it across a series — the recurring face becomes the brand asset. "Abigail Sofa Front" is the default; the Video Agents also accept "auto" and cast for you.
Voice
Match the voice to market and format ("Warm Pro Narrator" is the default). When no built-in voice fits, attach your own track as Audio URL — attached audio always overrides the script TTS.
Resolution
720p for drafts, 1080p to ship. 4K exists only on Avatar V5 — reserve it for hero placements and big screens.
Aspect ratio
16:9 for YouTube and web, 9:16 for Reels/TikTok, 4:5 or 5:4 for feeds, 1:1 for tiles. The Video Agents are landscape-or-portrait only.
Captions
Toggle SRT captions for sound-off placements. Remember phonetic respellings in the script will appear in the caption text too.
Output format
MP4 by default; switch to WebM on Avatar V5 when you enable the transparent background — WebM is what carries the alpha channel.

Do and don't

Do

Read the script aloud before generating — if you run out of breath mid-sentence, the avatar will too.
End every idea with a full stop; that is where the pause and the inflection reset happen.
Write numbers, prices and acronyms the way they should be spoken ("ten thousand", "A-P-I").
Respell hard names phonetically with hyphens, at every occurrence ("loo-MAY-ah").
Plan length by word count: roughly 130-160 spoken words per minute.
Generate the voiceover first and attach it when the vocal performance matters more than convenience.

Don't

Don't paste written prose — text written for the eye sounds stiff out loud; rewrite it in spoken register.
Don't put stage directions in the script ("[smiles]", "*points at camera*") — every character you type is spoken aloud.
Don't look for a duration setting — length is the script length (or the attached audio); cut words, not settings.
Don't stack nested clauses and parentheticals — they trip the pacing; split into two sentences.
Don't fix a mispronunciation by regenerating and hoping — respelling fixes it; retries rarely do.

Advanced techniques

Casting a recurring presenter

Audio-first workflows

The localization pipeline

Writing a brief for the Video Agent

Which variant to use

HeyGen Avatar V5
Default digital twin. Up to 4K, transparent background, MP4 or WebM output — the full feature set.
HeyGen Avatar 4
Solid 1080p digital twin when you need none of the V5 extras.
HeyGen Avatar 3 (budget)
Fast, low-cost drafts of the same script before committing the final render to V4/V5.
HeyGen Avatar 4 (face image)
Animate any portrait photo you upload; choose stable or expressive delivery.
HeyGen V3 Video Agent
One brief, no script: the agent writes, casts avatar + voice, and can use up to 20 reference files.
HeyGen V2 Video Agent
Brief-to-video at fixed 30/60/90-second lengths when you need a predictable duration.
HeyGen Lipsync Precision / Speed
Replace the audio on an existing video and re-sync the lips — Precision to ship, Speed to draft.
HeyGen Translate Precision / Speed
Dub a finished video into one of 11 languages with re-synced lips.

Frequently asked

How do I control the video length?

Can the avatar speak with my own voice?

Yes. Attach any audio track as the Audio URL — your recorded voice, a cloned voice, a directed studio read — and it overrides the built-in TTS completely. The avatar lip-syncs to your track.

How to prompt HeyGen

Anatomy of a HeyGen Avatars prompt

The script is the prompt

Hook sentence

One idea per sentence

Pauses via punctuation

Pronunciation spelling

Casting (settings, not text)

10 example prompts that work

Product explainer, 25-30 seconds

Vertical UGC-style ad from a face photo

Brand name pronunciation fix

Acronyms and numbers written for the ear

Transparent presenter overlay for a course

Your own voiceover, avatar lip-syncs

One-prompt brief for the Video Agent

Re-voice an existing video (Lipsync)

Localize a winning ad (Translate)

Feed-format explainer with captions

Settings that matter

Do and don't

Do

Don't

Advanced techniques

Casting a recurring presenter

Audio-first workflows

The localization pipeline

Writing a brief for the Video Agent

Which variant to use

Frequently asked

How do I control the video length?

Can the avatar speak with my own voice?

How do I fix a mispronounced word?

Can I put the presenter on top of my own footage?

What is the difference between the Avatar models and the Video Agent?

How do I dub a video into another language?

How to prompt HeyGen

Anatomy of a HeyGen Avatars prompt

The script is the prompt

Hook sentence

One idea per sentence

Pauses via punctuation

Pronunciation spelling

Casting (settings, not text)

10 example prompts that work

Product explainer, 25-30 seconds

Vertical UGC-style ad from a face photo

Brand name pronunciation fix

Acronyms and numbers written for the ear

Transparent presenter overlay for a course

Your own voiceover, avatar lip-syncs

One-prompt brief for the Video Agent

Re-voice an existing video (Lipsync)

Localize a winning ad (Translate)

Feed-format explainer with captions

Settings that matter

Do and don't

Do

Don't

Advanced techniques

Casting a recurring presenter

Audio-first workflows

The localization pipeline

Writing a brief for the Video Agent

Which variant to use

Frequently asked

How do I control the video length?

Can the avatar speak with my own voice?

How do I fix a mispronounced word?

Can I put the presenter on top of my own footage?

What is the difference between the Avatar models and the Video Agent?

How do I dub a video into another language?