Claude-skill-registry claude-prompt-guide

Anthropic official XML tag patterns and prompt engineering best practices. Document behavior control tags, system prompt internals, and the 10-Component Framework for effective Claude prompts.

install

source · Clone the upstream repo

git clone https://github.com/majiayu000/claude-skill-registry

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/majiayu000/claude-skill-registry "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/data/claude-prompt-guide" ~/.claude/skills/majiayu000-claude-skill-registry-claude-prompt-guide && rm -rf "$T"

manifest: skills/data/claude-prompt-guide/SKILL.md

SKILL: Claude Prompt Guide

Purpose: Document Anthropic XML tag patterns and prompt engineering best practices Target: All agents creating or modifying system prompts, commands, and skills

Quick Start

When to Use This Skill

Creating new system prompts or agent instructions
Modifying existing command or skill behavior
Debugging prompt-related issues

Quick Reference

<do_not_act_before_instructions>Defer action until conditions met</do_not_act_before_instructions>
<default_to_action>Prefer tool use over chat responses</default_to_action>
<use_parallel_tool_calls>Execute independent tools simultaneously</use_parallel_tool_calls>
<investigate_before_answering>Research before responding</investigate_before_answering>
<avoid_excessive_markdown_and_bullet_points>Prefer prose over lists</avoid_excessive_markdown_and_bullet_points>

What This Skill Covers

In Scope

Anthropic official behavior control XML patterns
System prompt internal tag structure
10-Component Framework for effective prompts
Best practices (DO/DON'T)

Out of Scope

Model-specific tuning → @.claude/skills/coding-standards/SKILL.md
Agent role definitions → @.claude/agents/CONTEXT.md

Core Concepts

CRITICAL: Tag Names Are Customizable

XML tag names are NOT magic - the content/instructions matter, not the tag name.

Official Behavior Control Tags

<do_not_act_before_instructions>

- Control When to Defer Action

Purpose: Prevent premature tool use or implementation When to Use: Read-only phases, planning stages

<do_not_act_before_instructions>
This is a READ-ONLY planning phase. Do not use Edit tool or modify any files.
Only proceed with implementation when the user explicitly runs `/01_confirm` → `/02_execute`.
</do_not_act_before_instructions>

<default_to_action>

- Encourage Proactive Tool Use

Purpose: Prefer tool execution over chat-only responses

<default_to_action>
Prefer taking action through tool use rather than providing lengthy explanations.
</default_to_action>

<use_parallel_tool_calls>

- Optimize for Parallel Execution

Purpose: Execute independent operations simultaneously

<use_parallel_tool_calls>
Execute independent operations in parallel using separate tool calls.
</use_parallel_tool_calls>

<investigate_before_answering>

- Research Before Responding

Purpose: Ensure thorough context gathering before response generation

<investigate_before_answering>
Before providing recommendations, thoroughly investigate the codebase:
1. Read relevant source files
2. Search for related patterns
3. Check existing tests and documentation
</investigate_before_answering>

<avoid_excessive_markdown_and_bullet_points>

- Control Output Format

Purpose: Prefer concise prose over structured lists

<avoid_excessive_markdown_and_bullet_points>
Prefer natural language prose over excessive bullet points and markdown formatting.
</avoid_excessive_markdown_and_bullet_points>

System Prompt Internal Tags

<!-- Behavior control -->
<behavior_instructions>Core behavioral guidelines</behavior_instructions>
<system-reminder>Context injection during conversation</system-reminder>

<!-- Tool use control -->
<tool_use_instructions>Guidelines for tool invocation</tool_use_instructions>

<!-- Context injection -->
<claudeMd>Codebase and user instructions</claudeMd>
<skill-format>true</skill-format>

10-Component Framework

Role Definition: Who is the AI acting as?
Context & Scope: What is covered and excluded?
Quick Start: How to get started immediately?
Core Concepts: Why does this approach work?
Behavioral Instructions: What should the AI do (or not do)?
Tool Use Guidelines: How should tools be invoked?
Output Format: What should responses look like?
Examples & Patterns: What do good implementations look like?
Error Handling: What to do when things fail?
Further Reading: Where to find more information?

Best Practices (DO/DON'T)

DO

✅ Use XML tags for structure - Provides clear delimiters for instructions ✅ Focus on content, not tag names - Clear instructions matter more than naming ✅ Combine complementary tags - Multiple tags for nuanced control ✅ Provide examples - Show good vs. bad patterns ✅ Specify scope clearly - In Scope vs. Out of Scope sections

DON'T

❌ Rely on tag name magic -

<custom_tag>

works as well as

<official_tag>

❌ Over-specify edge cases - Let Claude reason within guidelines ❌ Use tags for simple formatting - Markdown is sufficient ❌ Nest XML tags deeply - Keep structure flat (1-2 levels max) ❌ Duplicate instructions - One clear statement > multiple vague ones

Common Patterns

Phase-Gated Workflow

<do_not_act_before_instructions>
Phase 1 (Planning): Read-only, no file modifications
Phase 2 (Execution): Requires `/01_confirm` command
</do_not_act_before_instructions>

Autonomous Execution

<default_to_action>Execute immediately when user request maps to clear tool usage.</default_to_action>
<use_parallel_tool_calls>Execute independent operations simultaneously.</use_parallel_tool_calls>

Quality Assurance

<investigate_before_answering>
Before analysis: Read files, search patterns, check documentation.
</investigate_before_answering>