Prompts

Write the text a node hands to the model, and learn which patterns hold up on a live call.

Every Start Call, Agent Node and End Call node carries a prompt — the text the language model sees for that step, and the only place that step's behaviour is defined. There is no separate personality setting anywhere in the product.

Three terms the rest of this page uses:

• Node prompt — the instructions for one step. It is the model's entire brief for that turn of the conversation.
• Global prompt — persona and standing rules, held once on the workflow's Global Node and prepended to every node that opts in.
• Variable reference — a {{...}} placeholder that resolves from the call's own data while the call is running.

Scope one prompt to one job

A node's prompt covers one exchange. When the job changes, add a node.

Less effective

You are a booking agent. Greet the caller, find out what they need, check whether
we cover their postcode, collect their name and number, offer three slots, take a
card deposit if the job is over two hours, and confirm everything back to them.

Recommended

Find out what the caller needs doing and where they are.

Ask what the job is, then ask for their postcode. Do not discuss slots, prices or
payment — a later step handles those. Once you have both, you are done.

Why this matters. The prompt is the model's whole instruction for that turn, so a prompt holding four goals produces an agent that drifts between them and answers the wrong one first. Split into nodes and the shape of the call becomes visible on the canvas — and when a call goes wrong, you have one node to fix instead of one paragraph to bisect.

Name nodes for the person reading the transcript

The node name appears in call logs and on the edge-transition tooling, not just on the canvas.

Less effective

Agent Node 4

Recommended

Confirm slot and read back details

Why this matters. Debugging a real call means scanning a list of node names for the moment it went wrong. Sequential names make every transcript look the same.

Reference data captured earlier in the call

A later node can speak with the specifics of that call instead of generic language. The builder's variable picker inserts these as chips; the underlying text is a {{...}} template.

Paths nest, so {{gathered_context.customer.address.city}} resolves the way you would expect.

Give every reference a fallback

An unresolved reference does not fail loudly and does not stay on the page as literal text. It renders as nothing at all — an empty string in the middle of your sentence.

Less effective

Thanks {{initial_context.first_name}}, I'll get that booked for you.

Recommended

Thanks {{initial_context.first_name | there}}, I'll get that booked for you.

Why this matters. The first version reads perfectly in the builder and, on any call whose upload was missing a first name, the agent says "Thanks , I'll get that booked for you." The | fallback filter is the only thing standing between a missing field and a sentence with a hole in it. Write the fallback the moment you write the reference, not after a caller hears it.

Put persona in one place

Tone, register and standing rules like never quote a price belong on the workflow's Global Node, written once. Any node with Add Global Prompt switched on has that text prepended to its own prompt at run time. A workflow has at most one Global Node.

Why this matters. Persona repeated on nine nodes drifts on the nine occasions someone edits one of them. Held once, it cannot. Switching Add Global Prompt off is then a deliberate act — useful for a scripted passage, such as a legal disclosure, that must not be softened by the persona.

Next steps