How to Write AI Agent Instructions (System-Prompt Guide)

AI agent instructions are the agent's brain -- the plain-English brief that decides whether it is genuinely useful or quietly useless. Two people can give an agent the same apps and the same permissions and get wildly different results, and the difference is almost always here, in how well the instructions are written. This is the highest-impact part of building an agent, and it is the part most people rush.

On Deelo you write instructions as plain text in the AI Assistant -- no prompt-engineering background required, though a few habits make a large difference. This guide covers what to put in an agent's instructions, how to write each part so the agent actually follows it, and a reusable template you can paste and adapt. If you want the full builder flow around it, how to build a custom AI agent is the step-by-step; this guide is specifically about writing the brain.

You will not need every section at full length for every agent, but each one earns its place. The rest of this guide takes the parts that most affect behavior -- the role, the never-do list, the facts, the escalation triggers, and examples -- and shows how to write each so the agent follows it rather than interprets it loosely.

Start with a single sentence that names the job, the audience, and the point. 'You are a front-desk booking agent for a dental practice; you help patients schedule, reschedule, and confirm appointments.' That one line does more than it looks like -- it tells the agent what it is for, which shapes every judgment call that follows.

Resist the urge to make the role broad. 'You help with anything a customer needs' is not a role, it is an abdication, and it produces an agent that does everything poorly. A narrow role produces a focused agent. If you find yourself writing 'and also' more than once in the role sentence, you probably want two agents, not one.

The never-do list is the highest-value paragraph you will write, because it is where you prevent the failures that actually hurt. State the bright lines as absolutes, not suggestions. 'Never quote a price that is not on the official price list.' 'Never promise a delivery or completion date.' 'Never give medical, legal, or financial advice.' 'Always offer a human for anything involving a refund or a complaint.'

Write these as if the agent takes you literally, because it does -- that is a strength, if you are specific. A vague rule like 'be careful with sensitive topics' gives it nothing to act on; 'if a customer asks about anything other than booking, say you can only help with appointments and offer to connect them to the team' gives it a clear line. The instructions set intent; the permission grid enforces the hard limits underneath -- the two work together, so write the never-dos to match the guardrails you set.

An agent cannot answer from information it does not have, so put the real facts in the instructions. Your actual hours. The services you offer and what each includes. Your cancellation policy. The price list, if it should quote prices. Whatever the agent needs to be correct rather than plausible. An agent briefed with facts answers confidently and accurately; an agent briefed with adjectives improvises, and improvisation is where wrong answers come from.

There is a limit worth respecting: instructions are for the stable facts and rules, not for live data. The agent reads live data -- a specific customer's record, today's availability -- through its tools, like your CRM, not from its instructions. So put the durable stuff in the brief ('we're closed Sundays,' 'deposits are non-refundable within 24 hours') and let the tools supply the moment-to-moment specifics.

A great agent knows the edge of its competence, and you draw that edge with escalation triggers. Name the situations where it should stop and hand off rather than push forward: an upset customer, a request outside its scope, anything involving money or a dispute, anything it is genuinely unsure about. 'If you are not confident you have the right answer, say so and offer to connect a person' is one of the most valuable lines you can write.

This pairs with the permission grid, and the pairing matters. The instructions tell the agent when to defer; the grid enforces what it cannot do regardless of what it decides. Together they mean an agent that both chooses to escalate the tricky cases and is structurally unable to overstep on the dangerous ones. Write the triggers to match the risky verbs you gated -- if sends and refunds are on the approval floor, the instructions should tell the agent to route those to a human proactively, not wait to be blocked.

A couple of example exchanges teach an agent more than a paragraph of adjectives. Show it a model interaction: a customer asks something, and here is exactly how you want the agent to respond -- the tone, the length, the way it handles a curveball. Then show a hard case: here is a customer being difficult, and here is the graceful way to escalate.

Examples work because they are concrete where descriptions are abstract. 'Be friendly but efficient' is vague; a two-line sample of a friendly-but-efficient reply is unmistakable. You do not need many -- one good happy-path example and one good edge-case example cover most of what an agent needs to calibrate its behavior. Add them, and the agent stops guessing what you meant by your adjectives.

Paste that skeleton into the instructions field, fill each line with your specifics, and you have a brief that beats most agents in the wild. Then treat it as a living document: read the agent's runs, find the moments where it did something you did not want, and add a line that would have prevented it. Instructions are never done -- they get sharper every time you catch the agent guessing and replace the guess with a rule.