Skip to content

Give your agent a persona and instructions that stick

The difference between an agent people trust and one they abandon is its instructions — learn to write the persona, the rules, and the "when unsure" behavior that make it predictable.

Stage: Agent Builder · For: Maker · Level: Intermediate · Time: 15 min

When to use this

You've decided to build an agent, not just use one. The knowledge source is the easy part — Agent Builder will point at your docs. The part that decides whether the agent is any good is the instructions: who it is, how it answers, what it refuses, what it does when it doesn't know. A vague instruction block gives you a vague agent. This is the craft skill that separates a demo from something a team actually relies on.

This one is squarely for makers — it's the first real authoring muscle, and it transfers directly to Copilot Studio later.

What you'll need

  • M365 Copilot license with Agent Builder access
  • An agent in progress (or the intent to start one) and the knowledge source it answers from
  • A clear picture of the job: who asks it things, what "good" looks like, what it must never do

Try it now — the prompt

Instructions are the agent's operating manual. Write them like one — here's a strong starting block:

You are a support assistant for the [team] knowledge base. Answer only from the
connected docs. Keep answers under 5 sentences and always link the source. If the
answer isn't in the docs, say so plainly and point the user to the #[channel].
Never guess at policy, dates, or numbers. Use a warm, plain-spoken tone — no jargon.

Why this works: it sets the identity (who the agent is), the grounding rule (answer only from the docs), the format (short + cited), the failure behavior (say when unsure, route onward), and the tone — the five levers that make an agent behave the same way every time.

Step by step

  1. Open your agent and find the instructions field. This is the box that governs behavior — separate from the knowledge source you connect.
  2. Write identity, rules, and tone — in that order. Lead with who it is, then the hard rules (grounding, format, refusals), then the voice. Paste the block above and adapt it.
  3. Test the edges, not the easy path. Ask it something not in the docs and confirm it says so and routes you onward — that "graceful I-don't-know" is the whole game.
  4. Tighten the instruction that failed: 
    When you can't answer, don't apologize twice — say it once, give the channel
link, and stop. And never start an answer with "Certainly!"

Screenshots

Captured live in Microsoft 365 Copilot Agent Builder (Work mode). The product UI moves fast — if what you see differs, trust the numbered steps above, which we keep current.

Configure pane showing the agent name and a structured instructions block Identity first, then the hard rules, then the voice — the Instructions block reads like an operating manual, not a wish.

The instructions editor popped out into a focused dialog Pop the instructions out for a focused edit — grounding, citation, refusal, and tone in five tight lines.

Make it better

Instructions are never "done" — they're tuned: - Watch the transcripts. Every answer that's too long, too vague, or off-tone is an instruction you haven't written yet. Add a rule, re-test. - Encode the edge cases. When the agent does something surprising once, add the rule that prevents it forever — that's how a flaky agent becomes a reliable one. - Keep it tight. A wall of contradictory rules confuses the agent. Fewer, clearer instructions beat a sprawling list — edit for sharpness.

📚 Learn more. The Copilot Studio agent library overview shows how authored agents are structured, and the curated agent resources collect Microsoft guidance on building them well.

Watch out for

  • The agent follows instructions literally — so write literally. "Be helpful" means nothing to it. "Answer in under 5 sentences and link the source" means everything. Vague rules produce vague behavior.
  • Grounding is an instruction, not a guarantee. "Answer only from the docs" must be stated and tested. Don't assume it stays in scope — confirm it with an out-of-bounds question.
  • Tone is part of trust. An agent that's technically right but reads as cold or robotic gets abandoned. Spend a line on voice; it's not decoration.

Where this leads (the ramp)

You can now shape how an agent behaves — which is most of what makes one usable. The natural next questions are "how do I make it discoverable?" and "when do I outgrow this builder?" Those point straight at Stage 6 · Copilot Studio, where the same authoring instincts get pro-grade controls.

Next: Copilot Studio → Build your first Copilot Studio agent

Was this page helpful?

Found an issue or want to suggest an improvement? Open a GitHub issue  ·  Star on GitHub