A phone call has no scrollback. If your agent forgets the plan halfway through, the caller can't scroll up to re-read it — they just get confused and hang up. That single constraint is why a system prompt written for a phone agent needs a different shape than one written for a chat assistant.
Start with the call's one job
Before you write a single instruction, decide what a successful call looks like. "Book a repair appointment," "qualify a lead for the sales team," "answer FAQs and take a message otherwise" — pick one primary outcome per agent. An agent trying to do five things at once tends to do none of them well, because every instruction competes with every other instruction for attention.
The five-part structure we use
Every prompt we write follows roughly the same skeleton, in this order:
- Role & tone — who the agent is, and three adjectives for how it should sound
- Context it already has — business hours, services offered, anything it shouldn't have to ask
- The steps of the call — the exact order of questions or actions, written like a checklist
- Off-script handling — what to do when the caller asks something outside the flow
- Handoff conditions — the exact situations where the agent should transfer to a human
Writing the steps as a numbered checklist, rather than a paragraph of prose, makes a real difference. It gives the model a clear sense of where it is in the conversation, which shows up as fewer repeated questions and fewer dropped steps once the flow branches.
Three mistakes that quietly wreck call quality
- Writing the prompt as marketing copy instead of instructions — a paragraph about how great the business is doesn't tell the agent what to do next
- Leaving the fallback undefined — without a clear "if none of this matches, do X" rule, an off-script question can derail the whole call
- Testing only the happy path — the calls that break a flow are rarely the ones you scripted first
“If a step in your prompt can't be read aloud as a sentence a receptionist would actually say, rewrite it — it's a note to yourself, not an instruction to the agent.”
Test it before you trust it
Run a handful of simulated conversations against the draft before it ever answers a real call: the happy path, an impatient caller who interrupts, and someone asking something the flow doesn't cover. Simulation testing catches the gaps a quick read-through never will.
None of this needs to be perfect on day one. Publish, listen to the first real calls in the transcript log, and tighten the prompt around whatever actually confused a caller — that feedback loop matters more than getting the first draft exactly right.