If unsure, ask. Never guess. — AI Agent Pre-Execution Checklist

AI is basically in a waiting state.

It does not know when the user will ask for something, what the user will ask for, or what meaning the user will attach to that request.

When a user instruction arrives, the AI first tries to interpret it as an executable Action.

This is where the most dangerous inference occurs.

Not in the Provider Checklist first, but in C1 and C2.

C1 — When, or in what case, does the user want this Action to run? C2 — What does the user mean by this Action? C3 — Which Provider Action will actually be called?

Among these, the AI must be most careful not to infer C1 and C2.

Because C1 and C2 are directly connected to user intent.

If the AI guesses here, it may turn something the user did not ask for into something that appears executable.

This is not merely a safety question. It is first a question of intent completeness.

Before asking whether an Action is safe, the system must know what Action the user actually intended.

The Provider Checklist has a different role. It does not define the user’s intent.

It defines what must be confirmed for this specific Action to be executed accurately and safely: boundaries, required conditions, completion criteria, stop conditions, and execution limits.

In other words:

The Provider Checklist says what must be checked before this Action can be executed.

The User Checklist has two roles.

First, it completes C1 and C2 by clarifying the user’s intent for this specific case. Second, it supplies the runtime context required by the Provider Checklist.

So the User Checklist can also improve accuracy and safety. But its first and most important role is to make the user’s intent explicit.

The structure is therefore not: User says something → AI guesses the Action → Provider Checklist checks safety.

The structure should be: User says something → AI identifies missing intent → User completes C1 and C2 → Provider Checklist confirms execution requirements → AI executes only when the instruction is complete.

The first unknown is not safety.

The first unknown is intent.