Break the goal down into a plan of small steps and add a control point. Return JSON {"goal":"...","steps":["..."],"reviewPoint":"..."}: goal is the goal, steps is a list of sequential steps to be checked, reviewPoint is where you stop and check the intermediate result.
A big goal scares not only the person, but also the agent: it cannot be achieved in one step. Therefore, the goal is broken down into a sequence of small, verifiable steps - this is the plan. Mechanics: the agent works best when each step has a clear input and clear output, and the next step is based on the result of the previous one. The “plan first, action later” insider trick: Ask the agent to show you the step-by-step plan first and not execute it until you approve. This is how you catch the wrong logic before time is wasted. The second technique is control points: mark the places where you want to look at the intermediate result (after collecting data, before sending the letter). A typical mistake is to ask for everything at once in one giant step: “find, calculate, process and send.” If something goes wrong in the middle, you won't know where exactly and you'll have to start over. Break it down into steps, add a check point - and any failure will become local and quickly repaired.
Unlock access to submit solutions for instant AI review.