prompt-improvement

What this skill does

# Prompt Improvement

## MANDATORY: Query Official Documentation First

> **STOP - Before improving ANY prompt, you MUST invoke the docs-management skill.**
>
> This is NOT optional guidance - it is a required execution step.

### Required: Query Documentation via docs-management Skill

**Invoke the `claude-ecosystem:docs-management` skill BEFORE making any improvements:**

Search for relevant documentation using natural language:

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

### Why Documentation Query is Required

- **Training data is stale** - Your knowledge may be outdated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

### Recommended Query Topics

Query at least ONE of these topics based on your improvement needs:

- **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"

### Verification Checkpoint

Before improving a prompt, verify:

- [ ] 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?

**If ANY checkbox is unchecked, STOP and invoke the docs-management skill first.**

## Overview

This meta-skill replicates Anthropic's Console/Workbench prompt improver functionality within Claude Code. It provides workflows, patterns, and keyword registries to transform basic prompts into high-performance structured templates.

**What this skill provides:**

- 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)

**What this skill does NOT provide:**

- Duplicated official documentation (use `docs-management` skill)
- Hardcoded best practices (query `docs-management` for current guidance)

## When to Use This Skill

Use this skill when:

- **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

## Quick Decision Tree

**What do you want to do?**

1. **Improve a prompt from scratch** -> See [The 4-Step Improvement Workflow](#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](references/patterns/example-enrichment-patterns.md)
5. **Iterate with feedback** -> See [references/workflows/iterative-refinement.md](references/workflows/iterative-refinement.md)
6. **Generate test cases** -> See [references/workflows/test-case-generation.md](references/workflows/test-case-generation.md)
7. **Understand trade-offs** -> See [references/troubleshooting/tradeoffs-guide.md](references/troubleshooting/tradeoffs-guide.md)

## The 4-Step Improvement Workflow

The prompt improver enhances prompts through a structured 4-step process:

### Step 1: Example Identification

- 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)

### Step 2: Initial Draft

- 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"

### Step 3: Chain of Thought Refinement

- 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"

### Step 4: Example Enhancement

- 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 workflow:** [references/workflows/improvement-workflow.md](references/workflows/improvement-workflow.md)

## What You Get After Improvement

An improved prompt typically includes:

- **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

### Typical Improved Prompt Structure

```xml
<instructions>
Your task definition and behavioral guidelines
</instructions>

<context>
Background information and relevant details
</context>

<examples>
  <example>
    <input>Sample input</input>
    <thinking>Step-by-step reasoning</thinking>
    <output>Expected output</output>
  </example>
</examples>

<formatting>
Specify desired output format and structure
</formatting>
```

## Keyword Registry for docs-management

Use these keywords to query the `docs-management` skill for official documentation:

### Core Techniques

| Topic | Query Keywords |
| --- | --- |
| Prompt Improver | prompt improver, prompt improvement, optimize prompts |
| Chain of Thought | chain of thought, CoT, thinking, step-by-step, reasoning |
| XML Tags | XML tags, structure prompts, tagging, XML structure |
| Examples | multishot, few-shot, example formatting, multishot prompting |
| System Prompts | system prompt, role, persona, Claude role |
| Clarity | clear, direct, explicit, specific instructions, be clear and direct |
| Claude 4.x | Claude 4, Claude 4.5, Sonnet 4.5, Opus 4.5, best practices |

### Advanced Topics

| Topic | Query Keywords |
| --- | --- |
| Prefilling | prefill, response prefill, assistant prefill, output format |
| Long Context | long context, long context tips, document placement |
| Extended Thinking | extended thinking, thinking budget, deep reasoning |
| Prompt Chaining | chain prompts, prompt chaining, multi-step prompts |

### Example Queries

```text
Find documentation about chain of thought prompting and thinking tags
```

```text
Find documentation about XML tags for structuring prompts
```

```text
Find documentation about multishot prompting and example formatting
```

```text
Find documentation about Claude 4 and Claude 4.5 prompting best practices
```

```text
Find documentation about prefilling Claude's response for output control
```

## Performance Expectations

> **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.

Based on Anthropic's testing, prompt improvement typically yields:

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

**Imp