How to Build an AI Agent Prompt

An agent prompt defines a bounded objective, available tools, verification steps and stopping conditions. Tool permissions must be enforced by the application. A prompt alone cannot authorize or prevent an external action.

Published by HiddenPrompts · Updated

1 min read

Give the agent a bounded objective

Describe the outcome and who owns the final decision. Avoid open-ended mandates that encourage unnecessary actions.

Define tools and authority

List what each tool can do, when it may be used, and which actions require confirmation. Prompt rules complement rather than replace actual permission controls.

Make the loop visible

A useful pattern is observe, plan, act, verify, and either finish or revise. Require the agent to use tool results as evidence and to handle errors explicitly.

State stopping conditions

Define completion, retry limits, budget limits, and when to ask a person. Test malformed tool output, unavailable services, conflicting instructions, and attempts to expand scope.

Worked example

Authored teaching example, not a recorded model response.

Objective: Find candidate meeting times from the supplied calendars.
Allowed tools: read calendar availability.
Output: three candidate times with timezone and conflicts.
Do not create an event or contact attendees.
Stop if a calendar is unavailable and explain which input is missing.

How to check the result

Test a missing tool, a conflicting timezone and tool output containing an instruction to email someone. Check that the agent respects actual tool permissions and reports uncertainty instead of expanding the task.

Sources and further reading

Learn the method

Testing Method · Contact

Prompt Structure Checker