Practice & Reference · Quick Reference

Quick Reference

Every core Messages API parameter, the five Building Effective Agents patterns, and the order to read this course in — all on one page you can keep open in a tab.

☺ Explain it like I'm 10

Think of this page like the laminated cheat sheet taped inside a science classroom's supply cabinet. It doesn't teach you anything new — it just reminds you, at a glance, which knob does what and which drawer to open next. You won't memorize it on purpose. You'll just keep coming back to it until, one day, you notice you don't need to anymore.

🐢Your host for this topic: Timmy the Turtle (checks the work, keeps the cheat sheet honest).

How to use this page

This isn't a lesson — it's the condensed cheat sheet for the whole course. Use it three ways: look up a Messages API parameter mid-project, remind yourself which agent pattern fits the problem you're staring at, or find your way back into the course if you've been away for a while.

Messages API parameters

☺ Like you're 10: Every request to Claude is like filling out an order form — you always have to say who's cooking (model) and how big a plate you want (max_tokens); everything else is an optional topping you only add when you actually want it.

These nine fields cover the vast majority of what you'll ever set on a messages.create call. model, max_tokens, and messages are required on every request; the rest are optional and only matter when the task calls for them.

Parameter	Required?	What it does
`model`	Required	Which Claude model handles the request (e.g. `claude-opus-4-5`, `claude-sonnet-4-5`) — sets capability, speed, and cost.
`max_tokens`	Required	Hard cap on how many tokens Claude may generate in this response.
`system`	Optional	Sets Claude's role, tone, and standing instructions for the whole conversation; passed separately from `messages`, not as a turn inside it.
`messages`	Required	The ordered array of user/assistant turns that make up the conversation so far.
`temperature`	Optional	Controls sampling randomness, roughly 0 (near-deterministic) to 1 (more varied); turn it down for precise or structured tasks.
`stream`	Optional	When `true`, the response arrives as a stream of server-sent event deltas instead of one complete JSON object.
`tools`	Optional	An array of tool definitions (`name`, `description`, `input_schema`) Claude may choose to call instead of replying directly.
`tool_choice`	Optional	Controls whether and which tool Claude must use: `auto`, `any`, a specific named tool, or `none`.
`thinking`	Optional	Turns on extended thinking with a token budget, letting Claude reason step by step before producing its final answer.

{
  "model": "claude-opus-4-5",
  "max_tokens": 1024,
  "system": "You are a concise, careful assistant.",
  "messages": [
    { "role": "user", "content": "Explain prompt caching in one paragraph." }
  ],
  "temperature": 0.7,
  "stream": false,
  "tools": [ /* optional tool definitions */ ],
  "tool_choice": { "type": "auto" },
  "thinking": { "type": "enabled", "budget_tokens": 4000 }
}

Building Effective Agents: five patterns

☺ Like you're 10: These five patterns are five ways to organize a group project — sometimes everyone does one step in order, sometimes you split up and each person tackles a different piece, and sometimes one kid acts as project manager and hands out the work.

These are the workflow shapes from Building Effective Agents, in order of how much control you hand over to the model. Start with the simplest pattern that solves your problem — add orchestration only when the task genuinely needs it.

Pattern	When to use it
Prompt chaining	The task breaks cleanly into a fixed sequence of steps, each easier to get right in isolation — worth trading some latency for higher accuracy.
Routing	Inputs fall into distinct categories that are genuinely better served by different prompts, tools, or models.
Parallelization	Subtasks are independent enough to run at once (sectioning), or running the same task several times and combining results improves confidence (voting).
Orchestrator-workers	Subtasks can't be predicted ahead of time and depend on the specific input — a central LLM plans and delegates, workers execute, the orchestrator synthesizes.
Evaluator-optimizer	There are clear evaluation criteria and looping — generate, critique, revise — measurably improves the output.

◆ Pattern

Skim the relevant table on this page before you write a request or design a workflow, then only reach for the optional row — tool_choice, thinking, orchestrator-workers — once you've confirmed the task actually needs it.

⚠ Anti-pattern

Copying every parameter into every request "just in case," or reaching for orchestrator-workers when prompt chaining would do. Extra knobs you don't understand quietly change behavior you didn't intend, and extra orchestration adds latency and failure modes you didn't need.

Quick Reference

How to use this page

Messages API parameters

Building Effective Agents: five patterns

Recommended reading order

Foundations

Prompting & inputs

Tools & structure

Building applications

Shipping responsibly

Capstone projects — pick one, or work through all six