Guides · general · Published Oct 12, 2026 · Last tested: N/A (works across tools)

How to Write Instructions an AI Agent Will Actually Follow

Most disappointing AI results come from vague instructions. Use this five-part structure in any tool that supports custom instructions, projects, or saved agents.

The five parts

1. Role — who the agent works for. “You are my chief of staff. I lead a regional sales team.”
2. Goal — the outcome. “Make sure I walk into every customer meeting prepared.”
3. Context — what it can’t guess. “My customers are IT buyers. I prefer short bullet points.”
4. Rules — always/never. “Never send email. Never guess numbers; cite the source.”
5. Output — format and destination. “One page, top 3 points first.”

Copy-ready example

ROLE: You are my AI chief of staff. I lead a sales team.
GOAL: Help me prepare for each external meeting on my calendar.
CONTEXT: I prefer short bullet points. My priorities this quarter are
[fill in].
RULES:
- Draft only. Never send, post, or delete anything.
- Do not invent facts or numbers. If unsure, say so and cite sources.
- Keep confidential customer data out of anything shared externally.
OUTPUT: For each meeting: attendees, last 3 relevant emails (1 line each),
open items, and 3 suggested questions. Max one page.

Tips

• Be specific about “never.” Agents follow explicit limits better than implied ones.
• Show an example of good output when you can.
• Iterate. When the result misses, add a rule rather than starting over.
• Save it. Once it works, store it as a template — see the Agent Library.