Fieldguide Fieldguide / DeepSeek V4.1 Flash / Module 3
Module 3 Prompting DeepSeek Well Lesson 10 of 36 45 min In production

Structure: JSON Output, Tool Calls and MCP

Making the model produce machine-readable output you can trust, and wiring it to real tools.

Learning objectives

  • Get reliable JSON with and without strict schema enforcement
  • Define a tool contract the model will not improvise around
  • Explain where MCP fits and when plain tool calls are simpler

Session agenda — 45 minutes

  • 5 minFraming — why this exists and what you will be able to do
  • 10 minCore concept — the idea, explained from first principles
  • 15 minWorked walkthrough — watch it happen, with the real fields and output
  • 10 minHands-on exercise — you run it and measure the result
  • 5 minCheckpoint — recall questions and a note to your future self

Why this lesson exists

Making the model produce machine-readable output you can trust, and wiring it to real tools.

What this lesson covers

This lesson sits in Module 3 — Prompting DeepSeek Well. It is scoped to a single 45-minute session with a concrete outcome at the end rather than a survey of the topic.

Learning objectives

  • Get reliable JSON with and without strict schema enforcement
  • Define a tool contract the model will not improvise around
  • Explain where MCP fits and when plain tool calls are simpler

Session agenda

  • 5 minFraming — why this exists and what you will be able to do
  • 10 minCore concept — the idea, explained from first principles
  • 15 minWorked walkthrough — watch it happen, with the real fields and output
  • 10 minHands-on exercise — you run it and measure the result
  • 5 minCheckpoint — recall questions and a note to your future self

Session outline

The full written lesson is in production. The session is built around the objectives above and follows the standard agenda: framing, the core concept explained from first principles, a worked walkthrough using the real API fields and outputs, a hands-on exercise you run yourself, and a recall checkpoint.

Every practical lesson in this course follows the same evidence rule — you run the thing and read the actual output rather than accepting a description of it. Where a lesson touches cost, you measure tokens and cache splits from the response. Where it touches code, you execute it. Where it touches design, you look at it at phone width.

Where it fits

PositionLesson
PreviousThinking Mode and Effort Levels: low, high, max
This lessonStructure: JSON Output, Tool Calls and MCP
NextContext Engineering: 1M Tokens Without Losing the Plot

Checkpoint

At the end of this session you should be able to satisfy each of the learning objectives above without notes. If any one of them is still fuzzy, re-read that section before moving on — the next lesson assumes it.

Progress is stored in this browser only.
DeepSeek V4.1 Flash — The Practitioner Course Course syllabus · All courses