ElevenLabs Prompting Cheat Sheet

There is no prompt box. The script is the prompt — and v3 and v2 take incompatible syntax.

TL;DR — THE SHORT VERSION

There is no prompt field. The text you submit is the prompt.
v3 takes audio tags — [whispers], [excited], [sigh].
v3 does not support SSML break tags at all. A <break> in v3 text is simply not a pause.
v2 takes break tags — <break time="1.5s" />, 3 seconds maximum.
Punctuation is the universal control. Ellipses pause and add weight; CAPITALS emphasise.
A tag cannot override the voice. [shout] on a voice trained whispering will not shout.

VENDOR-DOCUMENTED

This sheet quotes the vendor's own published prompting guidance.

The formula

Script + punctuation + (version-appropriate tags)

Template:

[Sentence with ordinary punctuation for rhythm.] [Ellipsis for a weighted pause …] [CAPITALS for emphasis.] [v3 only: inline audio tag] / [v2 only: <break time="1.0s" />]

Worked example:

It was a VERY long day [sigh] … nobody listens anymore.

Breakdown — what each part does

The script

Ordinary prose with deliberate punctuation. This carries most of the performance.

Punctuation

Ellipses create a weighted pause. Standard punctuation sets rhythm. Works on every version.

Capitalisation

Increases emphasis on a word. The safest emphasis lever there is.

v3 audio tags

Inline square brackets for emotion and delivery. Effectiveness depends on the voice.

v2 break tags

<break time="x.xs" /> up to 3 seconds. Use sparingly.

Does it take a negative prompt?

Not applicable — there is no prompt or negative field.NO

The equivalent of a negative prompt here is restraint: fewer tags, not more. ElevenLabs warn that too many break tags in one generation "can cause instability" — the voice may speed up or introduce artefacts. Write the script plainly, generate, and only add a tag where the plain reading came out wrong.

What not to do

Do not use SSML break tags on v3 — they are not supported and fail silently.✕

Do not stack break tags. Too many in one generation causes instability and artefacts.✕

Do not expect a tag to contradict the voice's training.✕

Do not assume Professional Voice Clones behave the same on v3 — ElevenLabs state they are not fully optimised for it.✕

ORDER OF OPERATIONS

Write the script with punctuation first and generate. Add tags only where the plain reading failed. Tags applied pre-emptively to every line are the voice equivalent of a bloated negative prompt — they fight each other and destabilise the take.

The checklist

Before you send it:

1 — Know which version you are on — v3 or v2?□

2 — Using the tag syntax for that version only?□

3 — Punctuation doing the work before any tags?□

4 — Break tags used sparingly, under 3 seconds?□

5 — Tags tested on the actual voice, not assumed?□

SOURCE

ElevenLabs prompting documentation, elevenlabs.io/docs. Checked 25 Aug 2026.

Model versions and vendor documentation both move. Re-read the source before relying on a specific number. Errors are logged at corrections.

Every other sheet in this set: all model cheat sheets.