Prompting Cookbook

# HOW TO PROMPT CLAUDE & CLAUDE COWORK — COMPLETE REFERENCE

> **Purpose**: This document is a machine-readable reference file. Upload it to Claude so it understands Anthropic's official best practices for prompting itself and leveraging Cowork. All content is synthesized from Anthropic's documentation at docs.anthropic.com (February 2026). Non-coding focus.

---

## TABLE OF CONTENTS

1. [Core Prompting Philosophy](#1-core-prompting-philosophy)
2. [Be Clear, Direct, and Detailed](#2-be-clear-direct-and-detailed)
3. [Use Examples (Multishot Prompting)](#3-use-examples-multishot-prompting)
4. [Let Claude Think (Chain of Thought)](#4-let-claude-think-chain-of-thought)
5. [Use XML Tags to Structure Prompts](#5-use-xml-tags-to-structure-prompts)
6. [Give Claude a Role (System Prompts)](#6-give-claude-a-role-system-prompts)
7. [Chain Complex Prompts](#7-chain-complex-prompts)
8. [Long Context Tips](#8-long-context-tips)
9. [Extended Thinking & Adaptive Thinking](#9-extended-thinking--adaptive-thinking)
10. [Reduce Hallucinations](#10-reduce-hallucinations)
11. [Increase Output Consistency](#11-increase-output-consistency)
12. [Control Output Format & Verbosity](#12-control-output-format--verbosity)
13. [Claude 4.x / 4.5 / Opus 4.6 — Model-Specific Best Practices](#13-claude-4x--45--opus-46--model-specific-best-practices)
14. [Claude Cowork — Complete Guide](#14-claude-cowork--complete-guide)
15. [Prompt Templates & Patterns](#15-prompt-templates--patterns)
16. [Anti-Patterns to Avoid](#16-anti-patterns-to-avoid)

---

## 1. CORE PROMPTING PHILOSOPHY

Claude is a brilliant but context-free assistant. Each conversation starts with zero knowledge of user preferences, norms, styles, or workflows. Every prompt must supply all relevant context from scratch.

**The golden rule**: If a colleague with minimal context on your task would be confused by your instructions, Claude will be too.

**Prompting hierarchy** (most broadly effective → most specialized):

1. Be clear, direct, and detailed
2. Use examples (multishot)
3. Let Claude think (chain of thought)
4. Use XML tags for structure
5. Give Claude a role via system prompts
6. Chain complex prompts together
7. Optimize for long context
8. Use extended / adaptive thinking

Apply these techniques in order when troubleshooting. The actual impact depends on the use case.

---

## 2. BE CLEAR, DIRECT, AND DETAILED

### Principles

- **Provide contextual information**: Tell Claude what the output will be used for, who the audience is, where this task fits in a larger workflow, and what a successful result looks like.
- **Be specific**: If you want only a summary and nothing else, say so. If you want exactly 3 bullet points, state that.
- **Use sequential steps**: Numbered instructions ensure Claude follows the exact intended sequence.

### What to include in every prompt

| Element | Example |
|---|---|
| Task purpose | "This will be used in a board presentation" |
| Target audience | "Written for non-technical executives" |
| Workflow position | "This is step 2 of 5 in our content pipeline" |
| Success criteria | "A good result includes specific metrics and actionable recommendations" |
| Format requirements | "Respond in 3 paragraphs, no bullet points" |
| Constraints | "Stay under 200 words" |

### Add context to improve performance

Explaining WHY an instruction matters helps Claude generalize better than raw commands alone.

**Less effective:**