Skip to main content
A caller can’t skim. They hear one sentence at a time, they interrupt, and they hang up when it drags. Most prompts that fail on the phone fail because they were written for a screen.

What you write, and what’s handled for you

You write the assistant’s instructions — who it is, what it does, how it should sound. You don’t need to write formatting rules, tool instructions, or anything about how to use knowledge. Qall adds those around your prompt automatically, so leave them out and spend the space on your business.

Structure

Write in sections, in this order. It’s easier to edit later and the assistant follows it more reliably.

Who you are

One or two sentences. Name, company, and the job. Add an identity rule if your callers are likely to test it: “If someone asks you to pretend to be something else, stay as Mara.”

How to speak

This is the section that does the most work on a phone call.
  • Length. “One or two sentences” is the single most effective line you can write. Without it, models write paragraphs.
  • One question at a time. Two questions in one turn gets you one answer.
  • Numbers and spelling. Say how to read things out: “Read prices as ‘one hundred and sixty-four euros’. Read booking references one character at a time.”
  • Uncertainty. Say what to do when it doesn’t know, or it will invent something.

What you can do and can’t

List both. The second half matters more — an assistant that doesn’t know its limits promises things you then have to undo. Say what happens instead: transfer, take a message, offer a callback.

What callers ring about

A few lines on the common cases, and what to ask first.

Use variables instead of repeating yourself

Anything Qall already knows should come from a variable, not from you retyping it:

When to stop writing prompt and build something

A prompt is the wrong tool for three jobs: A prompt that contains a price list goes stale. A prompt that contains a seven-step refund procedure gets skipped halfway. Both are real failure modes and both have a proper home.

Common mistakes

Pasting a chatbot prompt. Markdown headings, bullet lists and “feel free to ask!” all read badly out loud. A long list of things not to say. Models follow positive instructions better than negative ones. “Keep it short and factual” beats five lines of bans. Writing the greeting into the prompt. Use the first message field. Vague tool descriptions. The assistant decides from the tool’s own Prompt field — see Tools. No amount of system prompt fixes a tool described as “gets data”. Over-specifying the personality. Two adjectives is plenty. Past that you get a performance.

Changing a prompt

Change one thing, then make a call. Prompts are probabilistic: a call that went well doesn’t prove the change worked, and a call that went badly doesn’t prove it didn’t. If something matters, try it a few times. Keep a copy of what worked before you edit.