STOP - Before improving ANY prompt, you MUST invoke the docs-management skill.

This is NOT optional guidance - it is a required execution step.

claude-ecosystem:docs-management

• Primary query: "prompt engineering chain of thought XML tags"
• Read the top results returned by the skill

• Training data is stale - Your knowledge may be outdated
• Official docs are canonical - Anthropic's current best practices
• Prevents hallucination - Ground improvements in real documentation
• Ensures accuracy - Latest Claude 4.x guidance

• Chain of thought: "chain of thought thinking tags reasoning"
• XML structure: "XML tags structure prompts formatting"
• Examples/multishot: "multishot prompting examples few-shot"
• Claude 4.x best practices: "Claude 4 prompting best practices"

• Did I invoke the docs-management skill? (not just read about it)
• Did I receive documentation from docs-management?
• Did I READ at least one official doc returned by the skill?
• Are my improvements based on what I read from official docs?

• The 4-step improvement workflow (quick reference)
• Keyword registry for docs-management queries
• Decision trees for improvement strategies
• Pattern libraries (XML tags, chain of thought, prefills)
• Before/after transformation examples
• Iterative refinement guidance
• Trade-off warnings (latency/cost implications)

• Duplicated official documentation (use

  docs-management

  skill)
• Hardcoded best practices (query

  docs-management

  for current guidance)

• Improving an existing prompt for better accuracy or structure
• Adding chain of thought reasoning to complex tasks
• Structuring prompts with XML tags for clarity and parseability
• Enhancing examples with reasoning steps
• Optimizing prompts for Claude 4.x models with explicit instructions
• Iterating on prompt quality with feedback loops
• Generating test cases when examples are lacking

1. Improve a prompt from scratch -> See The 4-Step Improvement Workflow
2. Add XML structure -> Query docs-management: "Find documentation about XML tags for structuring prompts"
3. Add chain of thought -> Query docs-management: "Find documentation about chain of thought prompting"
4. Add/improve examples -> See references/patterns/example-enrichment-patterns.md
5. Iterate with feedback -> See references/workflows/iterative-refinement.md
6. Generate test cases -> See references/workflows/test-case-generation.md
7. Understand trade-offs -> See references/troubleshooting/tradeoffs-guide.md

• Locate and extract any existing examples from the prompt
• Note the format and structure of current examples
• Identify if examples are missing (if so, consider generating them)

• Create a structured template with clear sections
• Add XML tags to separate components:

  • <instructions>

    - Task definition and behavioral guidelines

  • <context>

    - Background information and relevant details

  • <examples>

    - Demonstration cases

  • <formatting>

    - Desired output format specification
• Query docs-management: "Find documentation about XML tags for structuring prompts"

• Add step-by-step reasoning instructions
• Include thinking tags:

  <thinking>

  ,

  <analysis>

  ,

  <answer>

• Guide Claude through the problem-solving process
• Query docs-management: "Find documentation about chain of thought prompting"

• Update examples to demonstrate the new reasoning process
• Add

  <thinking>

  steps within examples showing intermediate reasoning
• Ensure examples match the output format specification
• This teaches Claude HOW to reason, not just WHAT to output

• Detailed chain-of-thought instructions guiding Claude's reasoning
• Clear XML tag organization separating components
• Standardized example formatting with step-by-step reasoning
• Strategic prefills guiding initial responses

Empirical Guidance: The metrics below are illustrative examples from Anthropic's testing at time of publication. Actual results vary by task, domain, and model version. These are NOT guarantees - use them as rough benchmarks for improvement potential.

• 30% accuracy increase on multi-label classification tasks
• 100% word count adherence on summarization tasks
• ~40% reduction in prompt iteration cycles

• Improvement Workflow - Detailed 4-step process with quality checks
• Iterative Refinement - Multi-pass improvement with feedback
• Test Case Generation - Creating examples when none exist

• Basic Transformations - Simple before/after examples
• Advanced Transformations - Complex multi-component prompts
• Domain-Specific - Code, analysis, creative, support domains

• XML Tagging Patterns - Complete tag library with usage
• Chain of Thought Patterns - Basic, guided, structured CoT
• Prefill Patterns - Strategic response prefilling
• System Prompt Patterns - Role/persona assignment
• Example Enrichment Patterns - Enhancing examples with reasoning

• Common Issues - Problems and fixes
• Debugging Guide - Systematic debugging
• Trade-offs Guide - When NOT to use improvement

• Keyword Registry - Complete keyword-to-topic mappings
• Tag Reference - Complete XML tag reference

• prompt-improver - Subagent that executes the 4-step improvement workflow
  • Auto-loads this skill for keyword registries and workflow guidance
  • Use when you need automated prompt improvement execution
  • Invoked by the

    /improve-prompt

    command

• /improve-prompt - User-facing command for prompt improvement
  • Input modes: direct text, file path, context, iterate
  • Supports

    --feedback

    for iterative refinement
  • Supports

    --generate-examples

    for prompts lacking examples

• claude-ecosystem:docs-management - Official documentation access (all documentation queries delegate here)
• claude-ecosystem:current-date - Get current UTC date for version history
• claude-ecosystem:skill-development - Creating and validating skills

• v1.0.0 (2025-12-03): Initial release - Meta-skill for prompt improvement with 4-step workflow, delegation to docs-management, comprehensive pattern library