Skip to content
Anthropic Prompting Best Practices
On Trove · from Anthropic

Anthropic Prompting Best Practices

Comprehensive guide to prompting best practices for Claude models (Opus 5, Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5).

Read a preview

You can read the contents and the first 2,000 characters.
Free accounts get the whole book.

  • By Anthropic
  • 14 chapters
  • Updated Nov 1, 2026
  • 122 readers

What’s inside

  • Introduction & Model Overview
  • Universal Prompting Techniques
  • Claude Opus 4.6 Best Practices
  • Chapter: Claude Opus 4.7 Best Practices
  • Chapter: Claude Opus 5 Best Practices
  • Claude Sonnet 4.6 Best Practices

What is Anthropic Prompting Best Practices?

Comprehensive guide to prompting best practices for Claude models (Opus 5, Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5), covering universal techniques, model-specific optimizations, extended thinking modes, and agentic patterns. Includes the full Opus 5 chapter (verbosity and narration control, written-deliverable length, task scope and over-verification, subagent spawning, self-correction, thinking-disabled artifacts) and the Opus 4.7 chapter, which now carries the Claude Code Operator's Rulebook — 16 RULE / WHY / APPLY triples plus a quick-reference card covering effort levels, adaptive thinking, session structure, tokenizer migration, and long-running autonomous work.

Read more

Anthropic · markdown · Listed Mar 5, 2026

  • best-practices
  • anthropic
  • prompting-guide

From the first page

Claude is a family of state-of-the-art large language models developed by Anthropic. All current Claude models support text and image input, text output, multilingual capabilities, and vision. Models are available via the Claude API, AWS Bedrock, and Google Vertex AI.

This chapter covers six models, each optimized for different use cases:

The most capable model, built for complex agentic coding and enterprise work, with particular strengths in long-horizon agentic tasks. Opus 5 completes full tasks rather than leaving stubs or placeholders, and it performs best when given the complete task specification up front and left to run. Thinking is on by default, and it ships with a 1M token context window as both the default and the maximum.

The most capable model, with particular strengths in long-horizon agentic work, knowledge work, vision, and memory. Opus 4.7 builds on the strengths of previous Opus models with improved reasoning depth and state tracking across extended sessions. It supports effort levels up to max and introduces the new xhigh effort level.

A highly intelligent model for building agents and coding. Opus 4.6 excels at long-horizon reasoning, complex multi-step tasks, and agentic workflows. It features adaptive thinking, 128K max output tokens, and supports effort levels up to max.

Preview · 2,000 of 158,376 characters

Building Your Agent Team: A Practitioner's Guide to Multi-Agent AI Systems
Google Gemini Prompting Best Practices
Anthropic Engineering Playbook: Agents, Harnesses & Infrastructure
Tanach — תנ"ך (Hebrew Bible)

Contents

  1. 01Introduction & Model Overviewp.1
    1. Overview of the Claude Model Familyp.1
    2. Model Comparison Tablep.1
    3. How to Choose the Right Modelp.1
  2. 02Universal Prompting Techniquesp.2
    1. Be Clear, Direct, and Detailedp.2
    2. Use Examples (Multishot Prompting)p.2
    3. Chain of Thought Promptingp.2
    4. Use XML Tags for Structurep.2
    5. Give Claude a Role (System Prompts)p.2
    6. Model Self-Knowledgep.2
    7. Chain Complex Prompts for Multi-Step Tasksp.2
    8. Long Context Tipsp.2
  3. 03Claude Opus 4.6 Best Practicesp.3
    1. Adaptive Thinking (Recommended)p.3
    2. The Effort Parameterp.3
    3. Long-Horizon Reasoning and State Trackingp.3
    4. Subagent Orchestrationp.3
    5. Balancing Autonomy and Safetyp.3
    6. Overthinking / Excessive Thoroughnessp.3
    7. Overeagerness / Overengineeringp.3
    8. Tool Usage Patternsp.3
    9. +7 more sections
  4. 04Chapter: Claude Opus 4.7 Best Practicesp.4
    1. Response Length and Verbosityp.4
    2. Calibrating Effort and Thinking Depthp.4
    3. Tool Use Triggeringp.4
    4. User-Facing Progress Updatesp.4
    5. More Literal Instruction Followingp.4
    6. Tone and Writing Stylep.4
    7. Controlling Subagent Spawningp.4
    8. Design and Frontend Defaultsp.4
    9. +4 more sections
  5. 05Chapter: Claude Opus 5 Best Practicesp.5
    1. Capability Improvementsp.5
    2. Calibrating Effortp.5
    3. Response Length and Verbosityp.5
    4. User-Facing Progress Updatesp.5
    5. Written Deliverable Lengthp.5
    6. Task Scope and Over-Verificationp.5
    7. Controlling Subagent Spawningp.5
    8. Self-Correctionp.5
    9. +3 more sections
  6. 06Claude Sonnet 4.6 Best Practicesp.6
    1. Default Effort Levelp.6
    2. Thinking Modesp.6
    3. Migration from Sonnet 4.5p.6
    4. Coding Use Casesp.6
    5. Chat and Non-Coding Use Casesp.6
    6. When to Try Adaptive Thinkingp.6
    7. When to Use Opus 4.6 Insteadp.6
    8. Breaking Changesp.6
  7. 07Claude Haiku 4.5 Best Practicesp.7
    1. Key Specificationsp.7
    2. Extended Thinking Supportp.7
    3. Best Use Casesp.7
    4. Cost Optimizationp.7
    5. Prompting Tipsp.7
    6. Haiku vs Sonnetp.7
  8. 08Extended Thinking Deep Divep.8
    1. How Extended Thinking Worksp.8
    2. Three Thinking Modesp.8
    3. Interleaved Thinking for Tool Usep.8
    4. Budget Optimizationp.8
    5. Prompting Tips for Extended Thinkingp.8
    6. Handling Redacted Thinking Blocksp.8
    7. Prompt Caching Considerationsp.8
    8. Thinking Block Preservation (Claude Opus 4.5+)p.8
    9. +2 more sections
See all 14 parts
  1. 09Agentic Patterns and Tool Usep.9
    1. How Tool Use Worksp.9
    2. Parallel Tool Calling Optimizationp.9
    3. Tool Usage and Triggeringp.9
    4. Reducing File Creation in Agentic Codingp.9
    5. Research and Information Gathering Patternsp.9
    6. Minimizing Hallucinations in Agentic Codingp.9
    7. Avoiding Hard-Coding and Test-Focused Solutionsp.9
    8. State Management Best Practicesp.9
    9. +3 more sections
  2. 10Chapter: Claude Haiku 5.5 Best Practicesp.10
    1. Key Specificationsp.10
    2. How Haiku 5.5 Differsp.10
    3. Use Effort to Control Thinkingp.10
    4. Accurate Search Resultsp.10
    5. JSON Output with Your Own Toolsp.10
    6. Prevent Early Stopping in Long Agent Promptsp.10
    7. Tell Coding Agents to Verify Their Changesp.10
    8. Mid-Turn User Messagesp.10
    9. +6 more sections
  3. 11Chapter: Claude Opus 5.5 Best Practicesp.11
    1. Capabilities Relevant to Promptingp.11
    2. Calibrate Effortp.11
    3. Prompts Written for Thinking Disabledp.11
    4. Unattended Agentic Runsp.11
    5. User-Facing Progress Updatesp.11
    6. Explore Context Before Acting (Multi-App Workflows)p.11
    7. Time Signals for Multiagent Harnessesp.11
    8. Thinking Instructions in Chat System Promptsp.11
    9. +7 more sections
  4. 12Chapter: Claude Sonnet 5.5 Best Practicesp.12
    1. Calibrate Effortp.12
    2. Steer Initiative and Scopep.12
    3. Running Without Up-Front Thinkingp.12
    4. Reasoning Tasks with JSON Outputp.12
    5. User-Facing Progress Updatesp.12
    6. Tool Use in Chat and Knowledge Workp.12
    7. Mid-Turn User Messagesp.12
    8. Verification on Coding Tasksp.12
    9. +5 more sections
  5. 13Chapter: Claude Fable 5.1 Best Practicesp.13
    1. Consider All Effort Levelsp.13
    2. Ask for User-Facing Progress Updatesp.13
    3. Batch Independent Tool Calls in Agent Loopsp.13
    4. Keep the Conversation History Append-Onlyp.13
    5. Writing Density and Formattingp.13
    6. Quoting Retrieved Sourcesp.13
    7. Finish the Whole Taskp.13
    8. Keep Changes and Tests to What the Task Asks Forp.13
    9. +3 more sections
  6. 14Chapter: Hooks, Skills and Subagent Prompts for the 5.5 Generationp.14
    1. What Changed in the 5.5 Generation That Affects Harness Textp.14
    2. Aggressive Language: MUST, CRITICAL, "If in Doubt"p.14
    3. Give the Reasonp.14
    4. Literal Instruction Followingp.14
    5. SessionStart Hook Textp.14
    6. Skill Descriptions and Bodiesp.14
    7. Subagent Prompts and Delegationp.14
    8. How to Make a Model Consult a Resource Before It Actsp.14
    9. +1 more section