Skip to content

PROMPTS.md

Latest commit

 

History

History
54 lines (39 loc) · 4.73 KB

PROMPTS.md

File metadata and controls

54 lines (39 loc) · 4.73 KB

Prompt Engineering Guidelines for Claude Code

The most effective prompt engineering techniques for Claude Code center around structured instructions, the use of sub-agents, emphasis styling (bold/uppercase), and optimal placement of instructions for code commands and workflows. These strategies give the highest-quality AI responses and reliable automation in Claude Code environments.[1][2][3][4]

Structured Prompt Writing

  • Use a clear, organized format with instructions separated from context, often via XML-style tags (e.g., <instructions>, <example>, <context>) to help Claude distinguish different prompt elements.[3][1]
  • Place primary instructions and context near the beginning of the .md command file or within the dedicated CLAUDE.md file in your project directory, as these are loaded first by Claude agents at session start.[5][6]

Bold and Uppercase Emphasis

  • For emphasis, Claude Code responds well to Markdown bold (**word**) for short, critical terms or instructions—especially where a decision or action is required.[1]
  • UPPERCASE is recognized and useful for acronyms, warnings, and strongly prioritized instructions, but should be used sparingly to avoid dilution of urgency or meaning.
  • Combine bold and uppercase only for highly important keywords (example: **WARNING: DO NOT DEPLOY**) for maximum model attention in generated responses.[1]

Claude Code Sub-Agents Best Practices

  • Sub-agents should have their own markdown/YAML frontmatter, including name, description, and (optionally) allowed tool lists, stored in .claude/agents/ for modular and discoverable workflows.[4]
  • Assign sub-agents explicit system prompts defining their role (e.g., "Implementer," "Tester," "Architect"), context window limits, and step outputs to achieve task separation and reproducibility.
  • Use hooks and chaining to pass outputs between agents; this ensures governance, separation of concerns, and parallelism, enhancing reliability across software pipelines.[6][4]

Instruction Location in Files

  • Place key instructions in CLAUDE.md for project-wide context and in each .md command or agent file for command-specific info; this enables Claude to load relevant information at runtime and support both individual and team workflows.[5]
  • You may also reference other files using @filename syntax to dynamically pull in contextual instructions or standards as needed.[5]

Iterative Refinement and Output Formatting

  • Request a specific output format (e.g., Markdown, JSON, table) in your prompt to control Claude Code's response structure.[2]
  • Iterate: Start with general instructions, review results, and refine progressively for complex automation tasks, leveraging chain-of-thought and chaining techniques to increase clarity and depth in reasoning.[3]

Summary Table

Technique Details Placement/Syntax
Structured Prompts XML/Markdown tags; chain-of-thought reasoning <instructions>, <context>, or top of .md file [1][3]
Bold/Uppercase Use Markdown for key terms, UPPERCASE for urgency **IMPORTANT**, WARNING
Sub-Agent Role Definition Distinct agents, YAML/Markdown frontmatter .claude/agents/ subdirectory
Hooks/Chaining Connect agent outputs via hooks, scheduler, permissions .claude/settings.json//hooks command [6]
Instruction Location CLAUDE.md for global, .md for specific commands Project root or relevant directory
Iterative Refinement Review-refine loop, conversation-based improvements Adjust sequentially in session

Applying these practices systematically gives you clearer, more reliable Claude Code output and ensures teams or solo developers reap consistent automation benefits.[7][2][4][6][3][1][5]

1 2 3 4 5 6 7 8 9 10