Beyond a Single Prompt

So far, every technique in this book has focused on a single prompt-response exchange. Many real tasks need more: a conversation that builds context over multiple turns, or a model that can take actions — calling tools, reading results, deciding what to do next — rather than just producing text.

Multi-Turn Conversations

In a multi-turn setup, earlier messages stay in context for later ones. This lets you break a complex task into steps: first ask the model to outline an approach, review the outline, then ask it to execute step one. Each turn benefits from the grounding established in previous turns, and you get a checkpoint to catch a wrong direction early instead of after a long single-shot generation.

Agentic Prompting

An agentic setup gives the model access to tools — search, code execution, file access — and lets it decide, based on the task, which tools to call and in what order, using each tool's result to inform its next step. The prompting principles don't change fundamentally: clear task framing, explicit format expectations, and good examples still matter. What's added is describing each tool's purpose and constraints clearly enough that the model chooses correctly, and designing tool outputs the model can actually parse and reason over.

Where This Adds Real Value

Tasks that require up-to-date information a static model can't know (agentic search)
Tasks that require precision a language model alone can't guarantee, like arithmetic or code execution (tool-delegated computation)
Long workflows that benefit from checkpoints and course-correction (multi-turn planning)

Closing Thought

Every technique in this book — zero/few-shot framing, chain-of-thought, system prompts, structured output, templates, and testing — still applies inside a multi-turn or agentic system. They don't get replaced by "more advanced" methods; they become the building blocks each individual step is made of.

Prompt Learning