Prompt Engineer

What it does

Prompt Engineer analyses your existing prompts and identifies structural weaknesses — vague instructions, missing context, absent constraints, and poor role framing — then produces improved rewrites with an explanation of each change. It can generate multiple prompt variants simultaneously, score them against a set of quality dimensions (specificity, clarity, constraint coverage, output format guidance), and run side-by-side A/B evaluations so you can objectively compare which variant performs best across your chosen model family.

The skill covers the full prompt design workflow from first draft to production-ready system prompt. It generates few-shot examples tailored to your task, builds chain-of-thought scaffolds that guide models through complex reasoning steps, applies XML tagging patterns optimised for Claude, and compresses verbose prompts to reduce token cost without sacrificing instruction fidelity. It works with Claude, GPT-4, Gemini, and Mistral, adjusting its recommendations to each model's documented behaviours and quirks.

Capabilities

• Chain-of-thought scaffolding — generates step-by-step reasoning frameworks that reduce model errors on complex tasks
• Few-shot example generation — produces high-quality input/output pairs matched to your task domain
• System prompt architecture — designs role, context, output format, and constraints sections from scratch
• Role definition — crafts precise persona descriptions that improve instruction-following consistency
• XML tagging for Claude — applies Anthropic-recommended structured tagging to separate instructions from content
• Prompt compression — removes redundancy and reduces token count while preserving semantic completeness
• Iterative debugging — identifies why a prompt produces unexpected output and suggests targeted fixes

How to install

skills add prompt-eng

Configuration

Add the following to your .claude/skills.json to configure model family, output format, and the number of prompt variants generated per run:

{
  "prompt-eng": {
    "model-family": "claude",
    "output-format": "markdown",
    "variants": 3
  }
}

Example

Send a weak prompt to the skill and it returns a fully restructured version with reasoning:

Improve this prompt: Summarize the document

Example output — improved prompt with structured sections:

## Improved Prompt (Variant 1 of 3)

**Role:** You are an expert analyst producing executive briefings.

**Context:**
<document>
{{DOCUMENT_CONTENT}}
</document>

**Task:** Summarize the document above. Think step by step:
1. Identify the document type and primary audience
2. Extract the 3-5 most important points
3. Note any decisions, deadlines, or action items

**Format:** Return a summary of 150-200 words using bullet points.
Do not include information not present in the document.

## Change log
- Added role to anchor tone and register
- Wrapped document in XML tags to separate content from instructions
- Added chain-of-thought steps to reduce hallucination risk
- Specified output length and format to constrain verbosity