STARTING A RUN

INSTRUCTION ASSEMBLY AND GOAL DEFINITION Once the goal object exists, the harness assembles the instructions the model will receive. Instruction assembly means combining several sources of guidance into one complete initial prompt. The sources usually include the standing system rules, meaning the permanent behavior the agent must follow. They also include the user's specific request, now in structured form. They often include project conventions, which are rules written by the team for a particular project, commonly stored in a file such as a conventions or rules file that sits with the project. They may include relevant facts about the environment, such as today's date, the user's time zone, or which tools are available. And they may include examples of good output, which show the model the desired shape of the answer. This is the same prompt composition you studied in Chapter 20, applied at startup. An important design question is conflict. What happens if the user's request conflicts with a standing rule? For example, the user asks the agent to email a customer list to a personal address, but the standing rule prohibits sending customer data outside the company. The harness needs a clear priority order. A common order places safety and policy rules above project conventions, project conventions above the user's convenience requests, and the user's request above the model's own preferences. When a real conflict appears, the right behavior is usually to refuse that part and explain, or to ask for approval from someone authorized. Goal definition is the careful statement of what success looks like. A vague goal is "improve the report." A defined goal is "shorten the report to two pages, keep all numbers unchanged, and keep the section headings." Defining the goal early has two big benefits. First, the model can aim at something specific. Second, the harness can later check whether the goal was met, which is the subject of Chapter 26. A good goal definition answers four questions: what should be produced, what must be true of it, what must not happen along the way, and how will we know it is done.