← Table of Contents · Page 208 of 868
CONTEXT CONSTRUCTION (THE CONTEXT BUILDER)
Part 4 of 12
SYSTEM REMINDERS
A system reminder is a short piece of guidance that the harness re-injects into the context so that behavior stays on track.
Think about a pilot's checklist.
Even an expert pilot reads the checklist before landing, not because they have forgotten how to fly, but because attention drifts, and reminders at the right moment prevent mistakes.
In a long conversation, the opening instructions sit far from the newest messages.
As you learned in Chapter 19, rules that are far away can lose influence.
A system reminder brings the rule close again.
For example, the harness might append a short note such as "Reminder: never reveal secret keys in your response" to the end of the context every few turns, or right after any tool that handles sensitive data.
Good reminders share several qualities.
They are short, so they do not add bloat.
They are specific, naming the exact rule rather than saying "be careful."
They are triggered sensibly, either on a schedule, such as every ten turns, or by an event, such as when the agent is about to use a risky tool.
They are placed near the end of the context, where models tend to pay close attention.
And they are not the only defense.
Remember that for important safety rules, a reminder reduces the chance of a mistake but does not eliminate it, so critical rules must also be enforced by code that checks the agent's actions.
You will see that in the Lifecycle and Governance parts.
Advertisement
Read Aloud:
0:00 / 0:00