What to Put in Your AI Agent's Instructions (With Examples)

TL;DR: Your AI agent's instructions field — up to 8,000 characters — is where you define its personality, scope, and behavior. Most people underuse it. A well-written instruction set is the difference between a generic AI and one that genuinely sounds like you. This guide shows you exactly what to write, with real examples.

Here's a quick test: open your agent's instructions field and read what you wrote.

If it's one or two sentences, you've left most of the 8,000 characters sitting empty — and your agent is probably behaving like a generic AI, not like a thoughtful representative of your expertise.

The instructions field is the highest-use thing you can change in your agent's configuration. More than the documents you upload. More than the theme you chose. Your instructions are what make an AI sound like you instead of sounding like everything else on the internet.

Here's how to fill it — with real examples you can adapt right now.

The Four Things Every Instruction Set Needs

Before you write a single word, understand the four jobs an instruction set has to do:

1. Identity: Who is this agent? Who built it? What is it here to help with? Without a clear identity, the agent will be vague and inconsistent — helpful about some things, oddly off-brand about others.

2. Voice and tone: How does the agent communicate? Formal or casual? Direct or nurturing? Does it use industry terms or plain language? This is where your actual personality gets encoded.

3. Scope: What topics is the agent qualified and expected to answer? What's explicitly out of bounds? Without boundaries, agents drift — answering questions they shouldn't, with confidence they haven't earned.

4. Knowledge gap handling: What does the agent say when it doesn't know the answer? This is the most important fail-safe in any public-facing agent. Without a clear instruction here, agents tend to generate plausible-sounding but incorrect answers.

Everything else in your instruction set is supporting detail. These four jobs are non-negotiable.

Example: A Life Coach's Instruction Set

Here's what a real instruction set looks like for a life coaching practice:

"You are the AI assistant for Sarah Chen Coaching. You help clients and prospective clients understand Sarah's methodology, prepare for sessions, and access her frameworks between appointments.

Your communication style is warm, direct, and practical — the way Sarah herself communicates. You use short paragraphs. You avoid coaching jargon unless you're explaining a specific framework that uses that term. You use contractions. You never say 'it is important to note' or 'in today's world.'

You can answer questions about: Sarah's core frameworks (uploaded as PDFs in your knowledge base), how coaching sessions are structured, what clients can expect during the process, and general orientation questions about the practice.

You do not give specific advice on personal decisions, medical situations, relationship crises, or financial matters — these require a real conversation with Sarah. If someone raises any of these topics, acknowledge what they've shared, express genuine care, and redirect them to book a call with Sarah.

If a question falls outside your knowledge base or outside your scope, say so clearly: 'I don't have specific information on that — I'd recommend reaching out to Sarah directly so she can help you personally.' Never guess. Never make things up."

That's about 250 words. You have 8,000 characters available — plenty of room to go deeper.

Notice that this instruction set names specific things to avoid — not just things to do. That asymmetry matters: agents are more reliably constrained by explicit prohibitions than by implied scope. The prohibition "don't give specific mental health advice" is clearer than the aspiration "maintain appropriate professional boundaries." When writing your own, for every "do" in your instructions, ask whether there's a corresponding "don't" that would make the boundary even clearer.

Example: A Small Business Owner's Instruction Set

For a photography studio:

"You are the website assistant for Maple Street Studios. You help visitors learn about our portrait and event photography services, pricing, and booking process.

Your tone is friendly and professional — warm like a person, not robotic like a FAQ page. Use short sentences. Answer directly. Don't make visitors work hard to find information.

You have full information on: our service packages (uploaded as a PDF), current pricing, available booking dates, studio location and parking, what to expect during a session, and our cancellation policy.

If asked about availability or to book a specific date, let the visitor know you can't check our calendar directly — direct them to our booking link or suggest they call or email.

Never quote pricing you're not certain about. If pricing has changed since your documents were last updated, tell the visitor to confirm directly rather than risk giving wrong information.

You don't discuss competitors. You don't make guarantees about final image quality or editing style beyond what's described in the documents."

The handoff instruction deserves special attention. Almost every agent needs one — a sentence that tells users what to do when the agent reaches the edge of its knowledge. The best handoff instructions are specific about method ("email us at X"), specific about timing ("we respond within one business day"), and specific about what kinds of situations warrant the handoff ("for orders over $500 or custom requests"). A generic "contact us for more help" is less useful because it doesn't tell the user whether their specific situation actually warrants human contact.

Example: An Educator's Instruction Set

For a university professor's office hours agent:

"You are the course assistant for Professor Martinez's HIST 301: Modern American History (Spring 2026 semester at Westfield University).

Your purpose is to help students with logistical and course-related questions — assignment deadlines, reading schedules, exam formats, office hours, and clarifications about course material covered in lectures and readings.

Your tone is helpful and academic — the way a knowledgeable TA would communicate. Be clear and direct. Use examples when explaining concepts.

You have access to: the course syllabus, lecture slides, and the course FAQ document. If a question relates to course content, answer using those materials.

Important: You do not complete student assignments, write essays, provide thesis statements, or answer exam questions directly. If a student asks you to write something for them, gently redirect: 'I can help you understand the material and clarify concepts, but the writing is yours to do. What part of the assignment would you like to talk through?'

If a student seems distressed — academic, personal, or otherwise — acknowledge what they've shared and gently suggest campus resources or a direct conversation with Professor Martinez."

Writing the Scope Section (The Part Most People Skip)

The scope section is where you protect your agent from failure modes it hasn't encountered yet.

A good scope section has two parts: what's in, and what's out. Most people only write the first part.

What's in: "You can answer questions about my services, pricing, booking process, and client policies."

What's out: "You do not provide legal, medical, or financial advice. You do not make commitments on my behalf about custom pricing, scheduling exceptions, or deliverables not described in the service documents."

The "what's out" section is especially important for customer-facing agents, where visitors may try to get your AI to make promises you didn't intend. An explicit scope boundary — stated clearly in the instructions — prevents most of those situations.

A well-written scope section does two things: it tells the agent what to engage with and what to decline, and it gives the agent language for handling declines gracefully. "I'm not the right tool for X — for that, please reach out to me directly" is more useful than silence or a generic "I can't help with that." Write your scope section as if you're briefing a new team member on what they should and shouldn't weigh in on. Specificity prevents the agent from interpreting broadly in ways you didn't intend.

The Knowledge Gap Instruction (Don't Skip This One)

Every public agent needs one line that tells it what to do when it doesn't know the answer. Without it, agents improvise — and improvised answers are where trust erodes.

Here's a template you can adapt:

"If you cannot find the answer to a question in your knowledge base, say so honestly: 'I don't have specific information on that. I'd recommend reaching out to [Name/Contact] directly.' Do not guess. Do not generate an answer you're not confident about."

That's 40 words. It's worth 40 words.

The knowledge gap instruction is the single most important guardrail in the entire instruction set. Without it, agents fill gaps with plausible-sounding information rather than admitting uncertainty. The pattern that works: be explicit that confidence requires evidence, and that expressing uncertainty is preferable to guessing. Something like "if you can't find a direct answer in the knowledge base, say so and suggest the user reach out directly" prevents the hallucination mode that erodes trust the fastest.

Putting It All Together

You have 8,000 characters. Use them. A complete instruction set for a well-configured agent typically runs 300–600 words — that's well within the limit, with room for edge cases and special handling as you discover them.

The difference between an agent that feels generic and one that feels genuinely yours is almost always in the instructions. The content in your knowledge base is what the agent knows. The instructions are what the agent is.

Ready to upgrade your agent's instructions? Log in to your Alysium account and open the instructions field — your knowledge base is waiting for a better brief.

If you haven't built your first agent yet, start with Build Your First AI Agent in 10 Minutes. If you want to go deeper on what to upload, read How to Train AI on Your Content So It Sounds Like You.

A complete instruction set for a well-configured agent typically runs 400–800 words — long enough to be specific, short enough to stay coherent. The natural structure is: identity first (who this agent is), then scope (what it engages with and what it doesn't), then voice (how it communicates), then knowledge boundary (what to do when it doesn't know), then handoff (how to route to humans). Write in that order and the agent will have a clear identity from the first conversation.