Skip to main content
The prompt is a single document in the centre column, and the only required content field. Everything goes there: who the agent is, how it opens, what it should achieve, and what it must not do.
Agent Prompt document with Context, Conversation Flow and Rules headings, the formatting toolbar above it, and a variables tray showing a client_first_name chip and a custom-field chip

One document, with the variables tray along the bottom

The toolbar gives you Bold, Italic, Heading and Bullet list, plus Insert Variable, Detect Variables and a find bar. Pasting from a document keeps its formatting.

​
How to structure it

​
The template

Use Template fills the document with a starter script under ## Context, ## Conversation Flow and ## Rules. If you’ve already written something, it asks Replace the current prompt? first. The template follows the interface language.
It’s a starting script, not a set of safeguards. Add your own rules.

​
What the model actually receives

Your document isn’t the whole prompt. On every turn the model is given, in this order:
  1. Your document.
  2. The FAQ, appended as fallback answers.
  3. A Customer Information block with the contact’s fields.
  4. The contact’s memory, if Use memory is on.
  5. The date and time, in the agent’s timezone.
  6. An SMS style block: short, plain text, no markdown, no emoji, full links.
Headings and bold in the document are for you. Replies go out as plain text.

​
The Language setting

The Language field does not force the reply language. Lumina always detects and mirrors the customer’s language. The setting only changes the language of the handoff summary and of calendar labels.To force a language, add a rule to the prompt: “Always reply in Spanish.”
The field’s help text is written for calls. On SMS it means what’s described above.

​
How a turn works

  • Inbound texts are debounced for 5 seconds: a burst of short messages gets one reply.
  • The model sees the last 20 turns of the conversation.
  • At most 3 tool calls per turn.
  • MMS attachments are stored (up to 10 per message), but the model only sees [The customer sent a <type> attachment].
  • A reply longer than 1600 characters is split into several sends, each billed. See What SMS costs.

​
Variables

The tray under the editor lists every variable the document uses. Clean unused drops any that no longer appear in the text.
Written as {{name}} and replaced with the contact’s real data: {{client_name}}, {{client_first_name}}, {{client_last_name}}, {{client_email}}, {{client_phone}} and {{client_company}}.Their aliases resolve too:
Every custom field on the contact resolves by its own name, case- and accent-insensitive: {{Cedula}} matches a field called “Cédula”.
Written as [[name]].
They aren’t substituted with any value. There’s nowhere to define their content — they reach the model exactly as written. Use them as notes to yourself, not as data.
Nine {{memory_...}} tokens that insert what the agent remembers about the contact. They only apply when memory is in Variables mode.See Contact memory.
Unlike WhatsApp, variables do resolve in the First Message and the Auto Close Message.

​
Creating variables

There are no preset chips. Either type the token straight into the document, or use Insert Variable → New variable… and choose Client Variable or Global Variable, a name, a colour and an optional description.
  • Detect Variables registers any {{…}} or [[…]] already in the text.
  • Clean unused removes chips no longer referenced.
  • Typing {{ in the editor opens an autocomplete of the agent’s variables.

​
FAQs

The FAQ button in the top bar opens FAQ Management. Press Add FAQ for each question and answer. They’re appended after the prompt, as fallback answers.
There is no knowledge base. You can’t upload documents or PDFs — FAQs are simply more prompt text. Keep them short.
FAQ Management window listing two questions with their answers and the Add FAQ button

FAQ Management: question and answer pairs appended after the prompt

​
What doesn’t exist

  • Temperature, length, style or batching controls — write the behaviour into the prompt.
  • A knowledge base or document upload.
  • A language enforcer — the Language field doesn’t do it; a prompt rule does.

​
Next step

Channel, routing and behaviour

Which number, who answers when, and how a session ends.