How to Write Instructions Your AI Agent Actually Follows
Jun 3, 2026 · 8 min read
Write agent instructions the way you'd brief a new hire on their first day: specific scenarios, exact phrasing for sensitive topics like pricing and refunds, and a clear list of what to hand off to a human rather than guess at. Vague, general instructions produce vague, generic-sounding responses.
Most first drafts of agent instructions read like a company handbook: broad, general, and written for a human who already knows the business. An agent doesn't have that context, so vague instructions produce vague answers.
The instructions that work best read more like on-the-job training for a new hire on day one — specific scenarios, exact wording for anything sensitive, and clear boundaries for when to stop and hand off to a person.
Pricing and refunds are the two places this matters most. Don't write 'be helpful about pricing' — write the actual numbers, the actual policy, and the actual sentence you want used when a customer pushes back.
Escalation rules deserve the same specificity. 'Escalate if needed' means nothing to an agent; 'escalate immediately if the customer mentions a safety issue, a legal threat, or asks for a manager by name' is something it can actually act on.
Treat the first two weeks after launch as a tuning period, not a finished product. Read real transcripts, find the moments the agent guessed instead of following an instruction, and tighten the instruction — not the agent's general behavior.
Let's tailor it to your business
We'll set everything up for a smooth launch — without the complexity.