Working with prompts

“Prompt” in Voxa means the message a Question or Message node sends to a contact. This guide is about writing them well — it is not about LLM prompt engineering. Voxa does not ship a built-in LLM; if you connect one (see LLM providers) the same principles apply inside the model’s system prompt.

Variable interpolation

Any stored variable can be dropped into a prompt body with double curly braces:

Thanks {{full_name}}! We'll send the voucher to the number you messaged us from.

Variables that don’t exist render as empty strings — they do not throw and they do not block the flow. That’s deliberate: an empty greeting is better than a crash. It also means you must defensively name variables — a typo fails silently.

Ask one thing at a time

WhatsApp is a texting surface. Walls of text get ignored; multi-part questions get half-answered. One prompt, one question, one variable:

✅  What's your full name?
✅  What city do you live in?

❌  What's your full name, city, and contact number?

Break compound questions into several Question nodes. Voxa’s flow builder makes this cheap.

Constrain where possible

Free-text input is the biggest source of data-quality problems. Where the universe of valid answers is small, say so:

Which size?
1. Small
2. Medium
3. Large

Then use a Validation node (or condition) to check the answer is one of 1, 2, 3, small, medium, large. Trim whitespace, lowercase, and normalise before storing.

Never promise what you can’t deliver

If your flow says “We’ll send your voucher in 5 minutes” — make sure the downstream system actually can. WhatsApp reads feel intimate; a broken promise lands harder than it would in email.

Handle the wrong answer gracefully

When a validation fails, reply with something short and specific:

Hmm, that doesn't look like a South African ID. Please try again.

Don’t say “Invalid input”. Don’t dump the regex. And give the contact a way out — either a retry limit (so they don’t loop forever) or an explicit exit keyword.

Use WhatsApp formatting sparingly

WhatsApp supports a small subset of markdown-ish formatting:

• *bold* renders bold.
• _italic_ renders italic.
• ~strikethrough~ renders strikethrough.
• Triple backticks render monospace.

Everything else is literal. No headings, no links with labels, no tables. Paste a raw URL and WhatsApp will auto-link and sometimes auto-preview it.

Opt-in and opt-out language

For any campaign that collects personal information, the first prompt should make consent explicit:

Hi! This is the 2026 Voxa launch competition. By replying, you agree
to our terms and privacy policy:
- https://voxa.software/terms
- https://voxa.software/privacy

Reply STOP at any time to exit.

Note

Automatic STOP-keyword handling is on the roadmap but not yet shipped. Until it is, treat STOP and UNSUBSCRIBE as commitments you implement via a dedicated condition in your flow.

Keep the tone consistent

Decide once, per tenant, whether you write like a friendly person or a brand. Mixing breaks the illusion and surfaces the automation. Example pair:

✅  Welcome! Drop your name below and we'll get you sorted.
✅  Welcome to the ACME Corp 2026 promotional campaign. Please provide your full name.

❌  Welcome to the ACME Corp 2026 promotional campaign. Drop your name below and we'll get you sorted.

Version your flows, not your prompts

Don’t create question_name_v2, question_name_v3 nodes inside the same flow. Voxa already gives you flow-level versioning — cut a new flow version when the copy changes materially.

Anti-patterns

• Greeting + ask in the same message (“Hi Sarah, what’s your email?”). Split them. Or don’t greet at all — the Message node before the first Question does that job.
• Over-personalising with uninferred variables. Don’t write "Welcome back, {{last_campaign_name}}!" unless you’re certain the variable is set.
• Asking confirmation after every step. One confirmation at the end of the flow is enough.