AIDealsPro is reader-supported. Some links on this page are affiliate links: if you buy through them we may earn a commission at no extra cost to you. It never changes our verdicts. How we review
An ElevenLabs agent system prompt has five parts: identity, the job, the knowledge it may quote, guardrails, and conversation style. The prompt itself bills no characters — but a chatty prompt makes long calls, and calls bill per minute: 15 free, 75 on Starter ($1 first month, shown until Nov 8 when checked). Checked 4 Nov 2026.
Best for: Anyone configuring an ElevenLabs voice agent — a tight system prompt is the highest-leverage cost control you own, because per-minute billing punishes every rambling call.
Not for: Anyone who wants the agent to improvise freely — guardrails and a knowledge base are the point; an unconstrained agent is a cost and compliance risk.
Starting price: $0 — system prompts are configured on every plan including free (15 agent minutes); Starter is $6/month ($1 first month, shown until Nov 8 when checked). Checked 4 Nov 2026.
What does an agent system prompt do?
The system prompt is the agent’s job description, personality, and rulebook in one text field: who it is, what it may do, what facts it may quote, and how it should speak. It sits beside the knowledge base (quotable facts) and guardrails (hard lines it must not cross) — this guide treats all three as one writing task.
Why it matters more here than in a text chatbot: every word the agent says extends a phone call, and calls bill per minute. A prompt that produces tight 20-second answers costs a fraction of one that produces wandering 2-minute answers. The prompt is free to write; its cost shows up entirely in call minutes.
The 5-part structure
Every effective agent system prompt has the same five parts, in this order:
1. Identity
Who the agent is: name, company, role. “You are Maya, the receptionist for BrightSmile Dental.” Without it, the agent introduces itself differently on every call, which confuses repeat callers and wastes opening seconds.
2. The job
The one task plus the definition of done. “Your job is to book appointments. A call succeeds when the caller has a confirmed date and time.” One job per agent — an agent asked to sell, support, and survey will do all three badly and at length.
3. Knowledge it may quote
The facts it may state — hours, prices, policies — with a rule to say “I don’t have that information” for anything else. Agents without it invent facts, and invented facts create the longest, most expensive calls: the correction call.
4. Guardrails
The hard lines: topics to refuse, offers never to make, data never to collect. “Never quote prices not in the knowledge base. Never offer discounts.” One line here prevents the call that costs you a customer.
5. Conversation style
How the agent speaks: brief, warm, one question at a time. “Keep answers under 20 seconds. Ask one question per turn.” This part directly controls cost — style instructions are minute instructions.
Example system prompts
Two complete examples using the 5-part structure. Adapt the bracketed parts; keep the shape.
Example 1: Dental receptionist
“You are Maya, receptionist for [BrightSmile Dental]. Book appointments; success = a confirmed date and time. Quote only the knowledge base — otherwise say you don’t have that information and offer a transfer. Never offer discounts or medical advice. Keep answers under 20 seconds, ask one question per turn, confirm bookings by repeating the date and time.”
Example 2: Order support
“You are Sam, support agent for [Northwind Outdoor]. Resolve order questions; success = the caller knows their order status or has a return label. Quote only the knowledge base. Never promise refunds beyond policy; never ask for full card numbers. Keep answers under 20 seconds, and if the issue needs a human, say so within the first minute and start a warm transfer with full context.”
Note the warm-transfer line in example 2: ElevenLabs supports warm transfers with full context, so the prompt should tell the agent when to use them rather than letting a stuck call burn minutes.
How the system prompt affects cost
The system prompt is free to write and edit — it bills zero characters. Its entire cost impact runs through call minutes, because agent billing is per minute from your plan’s included minutes: 15 free, 75 on Starter, 275 on Creator, 1,238 on Pro, with extra Speech Engine minutes at $0.08 ($0.160 burst).
The maths is unforgiving: 100 calls a month at 4 minutes each is 400 minutes (Creator plus extra), while the same 100 calls at 2 minutes is 200 minutes (inside Creator’s 275). Cutting one minute from the average call across 100 monthly calls saves 100 minutes — about $8 at the $0.08 extra-minute rate.
Three prompt choices drive most of the variance: answer-length instructions, question discipline (one per turn), and the escalation rule. Set a per-call minute cap alongside the prompt — the cap is the backstop for every prompt imperfection. For the plan numbers, see our agents cost guide.
Write a system prompt in 5 steps
Five steps, each with its cost. Writing is free — only test calls consume minutes.
- Step 1 — Write the one-sentence job. “Book appointments; success = confirmed date and time.” If you cannot write it in one sentence, split it into two agents. Cost: $0.
- Step 2 — Draft the five parts. Identity, job, knowledge discipline, guardrails, style — one or two sentences each, under 200 words total. Cost: $0.
- Step 3 — Add the cost controls. “Keep answers under 20 seconds. Ask one question per turn. Transfer within the first minute if the issue is outside the knowledge base.” These three lines are the budget. Cost: $0.
- Step 4 — Test five scenarios inside the free 15 minutes. Happy path, confused caller, angry caller, edge case, transfer — 2–3 minutes each. Edit the prompt between tests, not the voice. Cost: ~10–15 minutes of the free 15.
- Step 5 — Deploy with a per-call cap and review transcripts. The cap backstops the prompt; the longest calls always trace to one prompt line. Cost: per-minute billing — 15 free minutes, then plan minutes or $0.08/min.
5 rules for prompts that keep calls short
- One job per agent. Multi-job agents ramble; single-job agents finish.
- Under 200 words. A prompt longer than the average call script gets skimmed by the model.
- One question per turn. The highest-leverage sentence in any agent prompt.
- Escalate early. “Transfer within the first minute” beats a ten-minute stuck call.
- Never let it improvise facts. “Say you don’t know” is shorter — and cheaper — than any invented answer plus the correction call.
Common problems and fixes
The agent rambles and calls run long
The style section is missing or vague. Add “keep answers under 20 seconds” and “ask one question per turn,” then check the transcripts — the rambling almost always traces to one permissive line.
The agent invents prices and policies
The knowledge discipline is missing. Add “quote only the knowledge base; for anything else say you don’t have that information” — and fill the knowledge base with the actual documents.
The agent will not transfer a stuck caller
There is no escalation rule, so the agent keeps trying — on your minutes. Add: “if the issue is outside the knowledge base, say so within the first minute and start a warm transfer with full context.”
The prompt works in testing, fails in production
Test callers are polite; real callers are not. Add angry-caller and confused-caller scenarios to your tests, and add a guardrail for each failure in the first 20 production transcripts. Prompts are maintained, never finished.
Pros
- Free to write and revise — the only part of the agent with zero direct cost.
- Highest-leverage cost control: two sentences can cut minute burn by a third.
- One prompt serves every channel — phone, widget, WhatsApp — with no per-channel rewrite.
- Guardrails plus knowledge discipline make the agent auditable for compliance reviews.
Cons
- A loose prompt is a standing cost — every call pays for its verbosity at $0.08/min.
- Needs maintenance: the first 20 production transcripts will expose gaps no test caught.
- No prompt fixes a thin knowledge base — the agent cannot quote documents you never uploaded.
- Over-constraining backfires: 40 rules produce a hesitant agent that confirms everything.
- Style instructions are interpretive — “brief” means different things to different model runs.
Who should NOT over-engineer the prompt
If your agent handles one simple flow — “what are your hours?” — a 40-line prompt is worse than a 10-line one; every rule is a chance to confuse the model, and the call is 30 seconds either way. And if you have not run the five test scenarios yet, do not add rules for failures you have not seen — write the short version, test it, and let the transcripts tell you what is missing. For voice options to pair with the prompt, browse the voice library.
Frequently asked questions
What should an ElevenLabs agent system prompt include?
Five parts, in order: identity (name, company, role), the job plus its definition of done, knowledge discipline (quote only the knowledge base), guardrails (refusals, escalation), and conversation style (brief, one question per turn). Keep it under 200 words.
Does the system prompt cost credits?
No — the system prompt is configuration, not generated speech, so it bills zero characters. Its cost shows up entirely in call minutes: a chatty prompt makes long calls, and calls bill per minute (15 free, 75 on Starter, 275 on Creator, 1,238 on Pro, then $0.08/min).
How long should the system prompt be?
Under 200 words — one or two sentences per part, then stop. If it keeps growing past 200 words, the agent’s job is probably two jobs; split it into two agents instead of one long prompt.
How do I stop the agent from inventing facts?
Add knowledge discipline: “quote only the knowledge base; for anything else, say you don’t have that information and offer a transfer.” Then fill the knowledge base — the rule is useless without the documents behind it.
What is the difference between the system prompt and the knowledge base?
The system prompt sets behavior — who the agent is, what it does, its rules and style. The knowledge base holds the quotable facts — hours, prices, policies. Keep facts in the base where they can be updated without rewriting behavior.
How do I keep agent calls short?
Put three lines in the prompt: “keep answers under 20 seconds,” “ask one question per turn,” and “transfer within the first minute if the issue is outside the knowledge base.” Add a per-call minute cap as the backstop, then review transcripts weekly.
Sources checked
How agents work, guardrails, knowledge base, and warm transfers: official ElevenLabs voice agents page (elevenlabs.io/voice-agents), checked 4 Nov 2026. Per-minute pricing, included minutes, and concurrent calls: official ElevenAgents pricing (elevenlabs.io/pricing/agents), checked 4 Nov 2026. Extra-minute rates: official ElevenLabs API pricing (elevenlabs.io/pricing/api), checked 4 Nov 2026.
Last updated 4 Nov 2026. Prices checked on the official ElevenLabs pricing page on that date; confirm at checkout.