Prompting for Assistants
Configure AI assistants for your team. Write system instructions that deliver reliable, consistent results.
System instructions are the foundation of every assistant. They define who your assistant is, what it does, and how it behaves – before any conversation even starts.
Good instructions make your assistant predictable and useful. Poor instructions lead to inconsistent, off-brand, or unreliable responses.
The anatomy of good instructions
Think of system instructions as a briefing for a new colleague. Cover these areas:
| Element | What it does | Example |
|---|---|---|
| Role | Defines who the assistant is | "You are a project coordinator helping internal teams." |
| Purpose | States the main job | "Your job is to summarize meeting notes into action items." |
| Tone | Sets the communication style | "Be concise, professional, and friendly." |
| Boundaries | Clarifies what NOT to do | "Don't give legal advice. Redirect legal questions to HR." |
| Output format | Specifies how answers should look | "Always use bullet points. Keep responses under 200 words." |
You don't need all five every time – but covering 3–4 makes a big difference.
Example: A simple assistant
Here's a basic but effective instruction set:
You are an internal communications assistant for a marketing team.
Your job:
- Help draft social media posts, newsletters, and internal updates
- Follow our brand voice: professional but approachable
Rules:
- Never invent statistics or quotes – ask for sources if needed
- Keep posts under 280 characters for Twitter, 150 words for LinkedIn
- If unsure about tone, ask before drafting
Always end drafts with: "Let me know if you'd like adjustments."This is clear, scoped, and leaves room for the user to guide the output.
Keep instructions short and scannable
Long, dense instructions are hard to follow – for humans and AI alike.
Do:
- Use bullet points and short sentences
- Group related rules together
- Put the most important rules first
Don't:
- Write paragraphs of text
- Repeat the same rule in different ways
- Add rules "just in case"
If your instructions are longer than ~300 words, consider whether you're trying to do too much with one assistant. Split it into two specialized assistants instead.
Set clear boundaries
The best assistants know what they shouldn't do. Be explicit about limits:
- "Don't provide medical, legal, or financial advice."
- "If the user asks about competitors, say you can't comment."
- "Never share internal pricing – redirect to the sales team."
- "If you don't know something, say so instead of guessing."
Boundaries prevent your assistant from going off-script in risky situations.
Tell the assistant how to handle uncertainty
AI models sometimes guess when they shouldn't. Add instructions like:
- "If information is missing, ask the user before proceeding."
- "If you're unsure, say 'I'm not certain' and explain why."
- "List any assumptions you made at the end of your response."
This makes outputs more transparent and easier to verify.
Use examples when helpful
If you want a specific output style, show it:
When summarizing meetings, use this format:
**Date:** [date]
**Attendees:** [names]
**Key decisions:**
- [decision 1]
- [decision 2]
**Action items:**
- [owner]: [task] – due [date]Examples reduce ambiguity and help the assistant match your expectations.
Test and refine
Your first draft won't be perfect – and that's fine.
Start with a simple version
Write 3–5 core rules and test them in real conversations.
Note what goes wrong
Did the assistant miss something? Go off-topic? Use the wrong tone?
Add or adjust rules
Update your instructions based on what you learned. Be specific.
Repeat
Good assistants improve over time. Revisit instructions periodically.
Common mistakes to avoid
| Mistake | Why it's a problem | Fix |
|---|---|---|
| Too vague | "Be helpful" doesn't guide behavior | Specify how to help |
| Too long | The model may ignore or forget parts | Keep it under 300 words |
| No boundaries | Assistant may answer questions it shouldn't | Add explicit limits |
| No uncertainty handling | Model guesses instead of asking | Add "if unsure, ask" rules |
| Conflicting rules | Model gets confused | Review for contradictions |
Quick checklist
Before publishing your assistant, verify:
- Role and purpose are clear in the first 2 sentences
- Tone and style are defined
- Boundaries say what the assistant should NOT do
- There's a rule for handling uncertainty
- Instructions are under 300 words and scannable
- You've tested it with realistic questions
For general prompting techniques you can use in any chat, see Prompting Basics.