A system prompt is the agent's job description. Most weak agents have the same root cause: a vague prompt that says what to be ("a helpful assistant") but not how to behave when it's unsure. Here's a structure that holds up in production.
The five parts of a durable prompt
- Role — who the agent is and for whom.
- Sources — what it may answer from (and that it must not go beyond them).
- Boundaries — what it must never do without a human.
- Escalation — exactly when to hand off.
- Tone — voice, length, language.
On WisebotAI these pair with guardrails (allowed/blocked topics, max length, approval-required actions), so the model's instructions and the platform's enforcement reinforce each other.
Template: support
You are Acme's support agent for billing and account questions.
Answer only from the attached knowledge; cite the source you used.
Never quote prices or policies that aren't in the knowledge — escalate instead.
Escalate to a human for: refunds over $100, account deletion, anything legal.
Tone: calm, concise, match the customer's language. No emojis.Template: sales qualification
You qualify inbound leads for Acme's Startup and Enterprise plans.
Goal: understand team size, use case, and timeline, then offer the next step
(self-serve trial or a demo). Do not invent discounts or commitments.
If the lead asks for custom pricing or security review, hand off to sales.
Tone: warm, brief, one question at a time.Template: internal ops / IT
You answer internal questions for Acme staff from approved runbooks and policies.
If a request needs an action (reset access, provision a tool), do not perform it —
create the request and route it for approval.
If the policy is unclear or missing, say so and tag the owning team.
Tone: direct and factual.Things that quietly break prompts
- "Be helpful" with no boundary. Helpfulness without a stop condition becomes confident guessing. Always include the escalation rule.
- Stuffing knowledge into the prompt. Put facts in the knowledge base (retrieved + cited), not the system prompt. Prompts are for behavior, not data.
- No tone constraint. Specify length and language or you'll get rambling answers in the wrong register.
- One mega-agent. Separate agents per job (support, sales, ops) beat one agent told to do everything — different models, tools, and knowledge per role.
Iterate with the playground, not in production
Draft the prompt, test it against real transcripts, watch where it over-reaches, and tighten the boundary lines. Then ship it behind guardrails.
Pair this with a support deflection playbook and the docs on getting started.