ElevenLabs Prompts: 25 Voice and Direction Templates

TL;DR: ElevenLabs prompts are three different things wearing one name: API parameters with published ranges, text-level direction inside the script, and voice design prompts that describe a speaker rather than a performance. Audio tags such as [whispers] are documented for Eleven v3 and v3 Conversational only. Everything below is labelled by model.

What does prompting ElevenLabs actually mean?

It means three unrelated things, and conflating them is why so much advice on this topic fails silently.

Model parameters. These are numeric fields on the API request. They are typed, bounded, and documented. Stability, similarity boost, style, speaker boost, speed. They apply per request and override whatever is saved on the voice.

Text-level direction. This is anything you write inside the script itself to change how it is read: punctuation, capitalisation, ellipses, narrative dialogue tags, SSML break tags, phonetic transcription, and bracketed audio tags. Crucially, this layer is model-dependent in ways the layer above is not. The same six characters can be a documented control on one model and undocumented text on another.

Voice selection and design. This is not prompting a performance at all. It is choosing or creating a speaker. ElevenLabs is blunt about the priority order here. Its v3 guide opens with the sentence "The most important parameter for Eleven v3 is the voice you choose." and goes on to note that a whispering voice will not shout convincingly no matter what tag you attach. A voice design prompt and a script direction prompt look superficially similar and do completely different jobs.

Get the layers straight and most "why won't it do what I asked" problems resolve themselves. You were usually pulling a lever that does not exist on the model you selected.

Which ElevenLabs models exist right now?

Five text-to-speech models, plus a separate family for designing voices. Verified against the ElevenLabs models page on 27 August 2026.

Two model IDs that still circulate widely are now listed as deprecated: eleven_turbo_v2_5 and eleven_turbo_v2, both pointed at the equivalent Flash model. If a prompt pack tells you to use Turbo, it predates the current docs.

Worth knowing before you build against this: the default model_id on the Text to Speech endpoint is eleven_multilingual_v2, not v3. Omit the field and your carefully tagged script runs on a model whose documentation does not cover tags at all.

Which models accept audio tags, and which ignore them?

Audio tags are documented for Eleven v3 and Eleven v3 Conversational. Nothing else.

The clearest statement is on the help-center article about pauses, which splits its own answer by model: the tag section carries a heading reading "Audio tags (Eleven v3 only)", and the break-tag section is scoped to "Multilingual v2, Flash v2, and Flash v2.5". The v3 Conversational model card separately lists "Audio tags for fine-grained control" as a feature. So the honest matrix is two models in, three models out.

What you should not conclude is that a v2-family model strips brackets. ElevenLabs does not document that behaviour either way, and it does warn about the neighbouring case on its Text to Speech page: "Descriptive text will be spoken out by the model and must be manually trimmed or removed from the audio if desired." Treat brackets on Multilingual v2 or Flash as untested input. Generate one line, listen, then commit.

The tag syntax itself is looser than most guides admit. ElevenLabs states on its Text to Dialogue page that "Audio tags are natural-language instructions, not an enum parameter." There is no closed list to memorise. The published examples group into voice-related tags ([laughs], [whispers], [sighs], [sarcastic], [curious]), sound effects ([applause], [gunshot], [explosion]), and experimental ones such as [strong X accent] and [sings]. The Text to Dialogue page adds a third category the prompting guide does not: overall direction, with examples like [football] and [auctioneer].

Two caveats ElevenLabs publishes itself, and both matter more than the tag list. Tag effectiveness is bounded by the voice, and the docs are explicit about it: "Don't expect a whispering voice to suddenly shout with a" shout tag. And on Text to Dialogue, the tag feature carries a live disclaimer: "This feature is still under active development, actual results may vary."

What are the documented parameter ranges?

Here is where the layers separate cleanly. These are fields, not text, and they behave the same whatever you write in the script.

Two things in that table are worth stating plainly rather than burying.

The speed range is published twice, differently. The help-center article on pacing says "Possible values range from 0.7 to 1.2." and adds "This setting is available for all voices and all models." The live OpenAPI schema behind /v1/text-to-speech/{voice_id} types speed as a plain number with no minimum or maximum at all, while newer per-model schemas in the same spec do bound it to 0.7 and 1.2. Use the help-centre range. Do not build a slider on the assumption that the API will reject an out-of-range value, because the endpoint schema does not say it will.

Stability is not one setting across models. On v3, ElevenLabs calls it "the most important setting in v3" and exposes three presets rather than a slider: Creative, Natural and Robust. The trap is in the description of the last one, which the docs call "Highly stable, but less responsive to directional prompts but consistent, similar to v2." That is the sentence to remember. The setting that makes a v3 voice consistent is the same setting that makes your audio tags stop working. The guide recommends Creative or Natural when you want tags to land.

Text to Dialogue narrows things further. Its settings object in the live spec accepts stability and nothing else, with model_id defaulting to eleven_v3.

Which text conventions work on which model?

This is the part that generic prompt packs get wrong most often, because the conventions look interchangeable and are not.

Break tags are the documented pause mechanism on Multilingual v2, Flash v2 and Flash v2.5. The syntax is <break time="1.5s" />, ElevenLabs supports up to three seconds, and it warns that excessive use causes the model to speed up or introduce artefacts. On v3, this does not apply: "Eleven v3 does not support SSML break tags." Use ellipses, capitalisation and tags instead.

IPA transcription works natively on v3, wrapped in forward slashes, with no XML around it. ElevenLabs quantifies the reliability honestly, at "80-90% pronunciation consistency", and says it is not perfect.

SSML phoneme tags are narrower than almost anyone writes. The docs section is headed as being for v2 models, but the body narrows it in the next line: "Phoneme tags are only compatible with the" eleven_flash_v2 model. The same page separately notes that Multilingual v2 does not support them at all. For Multilingual v2 the documented fallback is alias tags and pronunciation dictionaries, of which you can attach up to three per request.

Punctuation and capitalisation are the one layer that carries across everything. Ellipses add pauses and weight, capitals add emphasis, and narrative dialogue tags steer emotion on every model. The cost, again, is that narrative direction gets spoken aloud unless you strip it in post.

Request stitching, the previous_text and next_text fields that keep prosody consistent across chunks, has its own exclusion: "Request stitching is not available for the" v3 model. That single line decides most long-form architecture. Narrate the book on Multilingual v2 with stitching; perform the characters on v3.

The 25 templates, labelled by model

Each block below is copy-paste. The label above it is not decoration.

Group A. Audio-tag direction (Eleven v3 and v3 Conversational only)

1. Emotional beat sheet, single speaker
[thoughtful] I've been going over the numbers since Tuesday.
[pause] And I keep landing in the same place.
[firmly] We're not shipping in March. [exhales] We're shipping when it's right.

2. Whisper-to-normal reveal
[whispers] Don't turn around. Not yet.
[normal] Okay. Now.
[excited] Look at what they built.

3. Sarcastic aside inside a straight read
The onboarding takes about four minutes.
[sarcastic] Assuming, of course, that nothing goes wrong.
[warmly] It usually doesn't.

4. Reaction, not a line
So I told him the deploy was already live.
[laughs] He went completely silent.
[laughs harder] For a full minute.

5. Accent switch mid-read
[curious] You want to hear something strange?
[strong French accent] "Zat is not how ze recipe goes, my friend."
[laughs] I have no idea where that came from.

6. Pacing without break tags
It was a VERY long day.
[sighs] Nobody listens anymore... nobody.
[quietly] Anyway. Tomorrow.

7. Support call de-escalation ladder
[professional] Thanks for calling, my name is Rae.
[sympathetic] That sounds genuinely frustrating, and I'm sorry.
[questioning] Can you tell me what the screen showed just before it froze?
[reassuring] Okay. That's a known one, and we can fix it right now.

8. Ad read: hook, body, close
[excited] Okay, this is the part nobody tells you.
[normal] Every prompt you write disappears the second you close the tab.
[thoughtful] Every single one.
[warmly] Unless you keep them somewhere.

Group B. Multi-speaker and dialogue (Eleven v3, via Text to Dialogue)

9. Two-hander with an interruption
Speaker 1: [cautiously] So I was thinking we could maybe—
Speaker 2: [jumping in] —ship it Friday?
Speaker 1: [surprised] How did you know that?

10. Overlap and cut-off
Speaker 1: [starting to speak] If we both talk at once—
Speaker 2: [overlapping] —it sounds like a real conversation.
Speaker 1: [pause] Sorry. Go ahead.

11. Three-speaker panel with reactions
Speaker 1: [confidently] The migration finished at 4am.
Speaker 2: [sceptically] All of it?
Speaker 3: [laughing] Ask him about the rollback.
Speaker 1: [deadpan] We do not discuss the rollback.

12. Narrator plus character
Speaker 1: [measured] She had read the letter four times before she understood it.
Speaker 2: [crying] "He's not coming back, is he."
Speaker 1: [gently] Nobody answered her.

Group C. Narrative direction, no tags (Multilingual v2, Flash v2, Flash v2.5)

13. Dialogue-tag emotion
"You're leaving?" she asked, her voice trembling.
"That's it," he said flatly. "That's the whole thing."
NOTE: delete the narrative tags from the audio in post, or they will be read aloud.

14. Timed pause script
"Give me one second to think about it." <break time="1.0s" /> "Yes. That works."
"Hold on." <break time="1.5s" /> "Say that again."

15. Brand-name pronunciation, Flash v2 only
Welcome to <phoneme alphabet="cmu-arpabet" ph="M AE1 D IH0 S AH0 N">Madison</phoneme> Analytics.

16. Brand-name pronunciation, Multilingual v2 (dictionary entry)
<lexeme>
  <grapheme>Claughton</grapheme>
  <alias>Cloffton</alias>
</lexeme>

17. Long-form chunk with stitching context
previous_text: "...and that brings us to the second half of the argument."
text: "The evidence here is thinner than the authors admit."
next_text: "Which raises an obvious objection, one they never address."

18. Number-heavy read
Send it to 0800 118 9007, reference AB-4471, before 5:30pm on 14/03/2027.
API: set apply_text_normalization to "on" for this request.

Group D. Voice design prompts (Text to Voice, not script direction)

19. Documented voice design format
Native English, southern British. Male, 45-55. Studio quality.
Persona: weary night-shift dispatcher. Emotion: dry, patient, unhurried.
Low-pitched and slightly gravelly, with deliberate pacing and long settled pauses.

20. Preview text that matches the prompt
"It's been quiet since about two. Quiet's not always good, mind you.
Quiet just means nobody's called it in yet."

21. Deliberately low-fidelity character
Native English, working-class New York. Female, 30-40. Poor audio quality,
sounds like a voicemail left in a stairwell. Persona: rushed, mid-argument.
Thin and slightly clipped, fast pacing with abrupt stops.

Group E. Parameter recipes (API request bodies)

22. Audiobook narrator, stable long-form
{
  "model_id": "eleven_multilingual_v2",
  "text": "Chapter one. The house had been empty for eleven years.",
  "voice_settings": {
    "stability": 0.75,
    "similarity_boost": 0.8,
    "style": 0,
    "use_speaker_boost": true,
    "speed": 0.95
  }
}

23. Expressive v3 take
{
  "model_id": "eleven_v3",
  "text": "[whispers] I never knew it could be this way. [normal] But I'm glad we're here.",
  "voice_settings": { "stability": 0.3 }
}

24. Two-speaker dialogue request
{
  "model_id": "eleven_v3",
  "inputs": [
    { "text": "[cautiously] Hello, is this seat-", "voice_id": "VOICE_A" },
    { "text": "[jumping in] Free? [cheerfully] Yes it is.", "voice_id": "VOICE_B" }
  ],
  "settings": { "stability": 0.4 }
}

25. Reproducibility harness
{
  "model_id": "eleven_multilingual_v2",
  "text": "The first move is what sets everything in motion.",
  "seed": 424242,
  "voice_settings": { "stability": 0.5, "similarity_boost": 0.75 }
}

Can you get the same take twice?

Mostly no, and ElevenLabs says so first.

Its Text to Speech page states that "The models are nondeterministic." and points you at the seed parameter for consistency while warning that subtle differences remain. The API reference is equally direct: seed is a best effort, and "Determinism is not guaranteed." The accepted range is 0 to 4294967295.

There is one interesting exception, and it is on a different endpoint. The Voice Design API documents its own seed field as a stronger guarantee: "Same seed with same inputs produces same voice." That asymmetry makes sense once you see the layers. Designing a voice is a one-time generation you want to be able to recreate. Performing a script is a sampling process that ElevenLabs has never claimed to pin down.

Practically, this means you should plan for regeneration rather than reproduction. ElevenLabs allows up to two free regenerations of identical content with identical parameters, which is the mechanism it expects you to use when a take has a glitch. If you need one canonical version of a line, render it, keep the file, and stop regenerating. The same logic applies across every generative surface, which is why we wrote a longer piece on temperature, seeds and determinism and on why you get a different answer every time.

What can a prompt not fix?

Three things, and the first two are the ones people discover late.

Consent is a product rule before it is a legal one. ElevenLabs publishes its own position rather than leaving it to interpretation, so cite it rather than paraphrasing it. Its voice-cloning FAQ states: "You can only create a Professional Voice Clone of your own voice. Even with their consent, you cannot clone someone else's voice." Its Prohibited Use Policy, last updated 17 August 2026, separately prohibits "creating or using ElevenLabs audio output to intentionally replicate the voice of another person" without consent or legal right, in a way that harms them, or in a manner intended to deceive listeners about whether the voice was AI-generated. The legal picture on voice likeness varies by jurisdiction and is genuinely unsettled. The platform rule is not unsettled, and it is stricter than most people assume.

Detectability is not zero, and it is moving. ElevenLabs ships an Audio Detector built on inaudible watermarking, with its older AI Speech Classifier as a fallback. As of the help-center page checked on 27 August 2026, watermarking covers all Text to Speech generations by free users plus select paid features, with a phased expansion described. The same page notes that "No ElevenLabs audio created prior to June 2026 carries a watermark." Generated speech is a strong draft for narration, prototyping, internal video, e-learning and character work. It is a poor choice anywhere the audio is expected to be a record of a real person saying a real thing.

Voice quality caps direction quality. The v3 guide is candid that Professional Voice Clones "are currently not fully optimized for Eleven v3", and recommends an instant clone or designed voice for v3 work during what it calls the research preview stage. No tag rescues a voice trained on flat, monotone samples. If your output sounds emotionally dead across every prompt you try, the fix is upstream of the prompt.

Where this leaves the tool, and us

The honest summary of ElevenLabs prompting is that there is less "prompting" here than the phrase implies, and more configuration. Two models take audio tags. Three take break tags. One takes IPA natively, one takes phoneme tags, and the most popular one takes neither. The parameters are typed and bounded and mostly do what they say. The single largest lever is a choice you make before you write a word, which is the voice.

Prompt Architects generates the prompt and the script direction. We do not generate audio, we are not a text-to-speech provider, and nothing on this page routes through us. If you want speech, ElevenLabs is where you go. What we are useful for is the part that decays fastest: keeping the model-specific version of a direction template saved and reusable, so that the Flash v2 pronunciation trick does not end up pasted into a v3 script six weeks from now. Our free plan covers 5 prompt enhancements per day, forever, according to the FAQ page at prompt-architects.com/faq, checked 27 August 2026.

Two closing habits worth more than any template here. Write the model ID at the top of every saved script, because the same text means different things on different models. And keep a scratch generation of one line before you commit a chapter, because the cheapest way to find out whether a tag lands on your voice is to listen to it once. For the adjacent craft of writing lines that survive being read aloud at all, why AI-written scripts flop covers the failure modes that no amount of delivery direction will save you from, and AI video sound design prompts covers the layer that sits underneath the voice.