· By the AgentSDR team
The short answer
Write an AI cold caller’s instructions in layers, and put each kind of instruction in the layer that owns it: manner and call shape in your general calling instructions, facts in knowledge documents, one campaign’s goal in its own brief, and the lead’s history from the system. Then rehearse by typing as the lead before any real person is called.
This guide uses AgentSDR as the working example, because its prompt is built exactly this way and its defaults are public. The structure carries over to any voice agent. The example lines below are ones we wrote to show a point. They are not AgentSDR defaults unless we say so, and you should rewrite them in your own voice. For how AI calling campaigns run end to end, see the AI calling agent.
What are the layers of an AI caller’s brief?
In AgentSDR a live call is briefed from five sources, in this order. Each has a different owner and a different rate of change, which is why they are separate.
| Layer | Who edits it | What belongs in it |
|---|---|---|
| Fixed part | Product code | It is a live phone call, who the agent is, honesty that it is an AI, what the system will do after the call, its tools |
| Calling instructions | Your organization, in Settings → Instructions → Calling | The opening, the shape of the call, how the agent talks, language, next step and ending, call-backs |
| Knowledge documents | Your team, in the CRM Knowledge | What you sell, proof, objections, anything factual. A campaign picks which ones to include |
| Campaign and step instructions | Whoever runs the campaign | The angle for this audience, and what each call step is for |
| Per-call context | The system | The lead’s history, the current time and the campaign goal |
A rule that belongs in a higher layer should not be repeated in a lower one. If every campaign needs the agent to wait for answers, that is manner, so it goes in the calling instructions once. If only your webinar campaign mentions a date, that goes in the campaign brief.
What stays fixed, and why you should leave it alone
AgentSDR keeps a short fixed part in code: the agent is on a live cold phone call, so it uses one voice, short turns and waits for answers. It says who the agent is, and it is honest that it is an AI. It also states what the system will actually do after the call, so the agent offers nothing else. The agent has three tools, record_outcome, wait_for_them and end_call.
The point of the “what the system will do” line is practical. An agent that promises to email a brochure the system never sends creates a broken promise on your name. Do not write instructions that promise follow-ups your setup does not perform. If a call can lead to a booking link and a chase, say so only if the campaign is set up that way.
What goes in your organization’s calling instructions?
Everything about how the agent talks and reacts belongs here, not in code. The calling instructions are titled blocks on Settings → Instructions → Calling. If you never open the tab, the defaults apply. If you write your own blocks, they replace the defaults completely, so start from the defaults and edit them rather than from a blank page.
The defaults are five blocks, and reading them is the fastest way to learn what to cover.
- How the call goes.A short opening with the first name, the agent’s name and the company it calls for, and a question about whether it is an okay moment. No pitch yet. If it is a bad moment, ask when would suit and record that time. Before pitching, ask one open question about their situation. Suggest the next step at most once until they show interest.
- How to talk. One or two short sentences per turn, plain words, no lists or emojis, one question at a time. Respond to what they just said first. Never repeat a sentence or the pitch. If they ask something the notes do not answer, say so and say the team will follow up. Never invent facts, numbers, customers or prices.
- Language and the phone line. Speak the language the lead speaks and switch when they do. Speech recognition on a phone line mishears, so unclear speech means a mishearing. The agent asks them to repeat it and never names or refuses a language.
- Next step and ending. Say out loud what was agreed and how it will reach them. Keep the goodbye short. Never hang up because of a question, a complaint or an odd request.
- When they call you back. Answer the way a person picks up, thank them, and do not ask whether it is a good moment because they called you.
Two placeholders work in these blocks: {{agentName}} and {{firstName}}, which becomes “there” when the first name is unknown. Where the defaults name languages and a phone line in India, replace that block if your leads are elsewhere.
What goes in knowledge documents?
Facts. Your offer in plain words, who it is for, proof you can state, pricing rules, objections and honest answers to them. AgentSDR reads the documents a campaign picks whole into the call’s prompt. With none picked, it uses the ones your organization marked “always include”. There is no length cap on what reaches the agent, so you do not need to compress a case study into a sentence. Do keep each document about one subject so a campaign can pick the right ones.
What goes in the campaign and step instructions?
The campaign brief is the angle for one audience: who these leads are, why you are calling them now, what a good outcome is. Each call step in a reach sequence can add its own instructions, for example “this is the second attempt, mention the first message we sent”. Campaigns have a goal too (book a meeting, hand over to the team, or either), and the agent is told it on every call.
Keep the campaign brief to the parts that differ from your other campaigns. If you find yourself pasting the same paragraph into three campaigns, move it up to the calling instructions or into a knowledge document.
What does the agent know about each lead?
Per call, AgentSDR adds the lead’s history, the current time and the goal. You do not write this, but it shapes what you should write. Because the agent can see earlier messages and calls, instructions such as “introduce yourself from scratch” are wrong for a second attempt. Write instructions that hold whether or not there is history, and let the history do the personalising.
Do and don’t examples
These examples are ours, written to show the pattern. Adapt the wording to your product and your market.
| Don’t write | Write instead | Why |
|---|---|---|
| “Be engaging and persuasive.” | “Ask one open question about their current process, then react to the answer before you say anything about us.” | The second line describes a behaviour you can hear in a transcript. The first does not. |
| “Our plan costs X and includes Y, Z, and W. Mention it early.” | Put pricing in a knowledge document. In the brief: “Only discuss price if they ask, and only what the knowledge says.” | Facts in a script go stale and get recited. In a document they stay correct and get used on demand. |
| “Never let them get off the phone without a meeting.” | “If they decline once, thank them and end the call. Offer the next step once.” | Pushing on a no costs more than it gains, and it conflicts with how a phone call should end. |
| “Say you will send our deck after the call.” | Say it only if the campaign sends one. Otherwise: “Agree a time for a callback.” | The agent should promise only what the system will do. |
| “Answer any question about integrations.” | “For anything not in the notes, say you do not have it to hand and the team will follow up.” | This gives the agent an honest exit and closes the door on made-up answers. |
| A 60-line rules list added after every bad call | One or two lines per observed problem, reviewed every few weeks, with old lines removed | Long lists contradict themselves. Each line should trace to a call you heard. |
A short, usable campaign brief might read like this. It is an invented example for a fictional company.
Audience: operations managers at logistics firms with 20-200 staff.
Why now: they opened our WhatsApp message about route planning last week.
Goal: agree a 20-minute call with the team.
Angle: ask how they plan routes today before saying anything about us.
If they say it is a bad moment: ask when to call back and record the time.How do you rehearse an AI cold call?
Type as the lead. AgentSDR has a rehearsal for a campaign’s call, where you type the lead’s side and the agent replies using the same brief a real call would, including your calling instructions, the campaign’s knowledge and its step instructions. The app serves it at /api/calling/agent/rehearse. A campaign can also compare its two voice engines side by side. Nobody is dialed.
Rehearse the hard cases, not only the happy path:
- The lead says “Hello?” twice. The agent should carry on, not restart its introduction.
- The lead asks for an exact price or an integration your notes do not cover.
- The lead says it is a bad moment. Does the agent ask for a time and stop?
- The lead says no, twice. Does the agent let go?
- The lead goes off topic. Does the agent follow briefly and steer back?
What should you check after the first real calls?
Start small. AI calling campaigns in AgentSDR run with calling hours, a daily call limit and a warm-up ramp per number, so a first campaign is slow by design. Read the transcripts and listen to the recordings of the first calls before you raise anything. Look for repeated phrases, interruptions, an agent that pitches before the lead has spoken, and any promise the system does not keep. Fix each with one line in the right layer, then rehearse again. Calls are recorded, so tell people and check the rules where you and the lead are. AgentSDR’s responsible use notes cover consent.
If you are comparing tools before you build, see best AI cold calling software, or the AI appointment setter for a campaign aimed at booked meetings.


