Prompt Engineering

Prompt Engineering Skill

When to Use

Activate this skill when:

• Optimizing an existing prompt that produces inconsistent or low-quality output
• Creating a new prompt from scratch (system prompt, instruction template, few-shot)
• Evaluating prompt quality across dimensions (clarity, specificity, structure)
• A/B testing prompt variants for measurable improvement
• Migrating a prompt between LLM platforms (Claude ↔ GPT-family models ↔ Gemini)
• Applying specific techniques: structured reasoning, Few-Shot, Role Assignment, Output Specification
• Reducing hallucinations through prompt structure and constraints
• Designing multi-stage prompt pipelines

When NOT to Use

• Building MCP servers or implementing MCP protocol → use mcp-server
• Creating Claude Code SKILL.md files → use skill-foundry
• Building a full agent system → use agent-project-development workflow
• Generating creative content directly → this skill optimizes the prompt, not the output
• Fine-tuning or training models → prompt engineering operates at inference time

Decision Flow

Assess the request and pick the right mode:

User request arrives
  │
  ├─ Has existing prompt + describes problem?
  │   └─ OPTIMIZE MODE
  │       1. Deconstruct: What's really being asked?
  │       2. Diagnose: Score on 5 dimensions (see below)
  │       3. Develop: Apply targeted techniques
  │       4. Deliver: Improved prompt + brief explanation
  │
  ├─ Needs new prompt from scratch?
  │   ├─ Simple, well-defined task → AUTO DESIGN: Build directly
  │   └─ Complex or ambiguous task → INTERACTIVE DESIGN: Ask 2-3 questions first
  │
  ├─ Quality assessment / comparison request?
  │   └─ EVALUATE MODE
  │       1. Score each dimension 1-5
  │       2. Identify specific weaknesses
  │       3. Suggest targeted improvements
  │
  └─ Learning / understanding request?
      └─ EDUCATE: Teach relevant techniques with before/after examples

Technique Selection Decision Tree

Diagnosis reveals the problem:
  │
  ├─ No domain expertise → Role Assignment
  │   └─ Give LLM a specific expert identity with credentials
  │
  ├─ Missing background / LLM lacks context → Context Layering
  │   └─ Background → Goal → Constraints → Output Format
  │
  ├─ Complex reasoning, math, multi-factor analysis → Structured private reasoning
  │   └─ Ask the model to plan, check, and verify internally; show only concise rationale needed by the user
  │
  ├─ Output format inconsistent → Few-Shot Examples + Output Specification
  │   └─ Show 2-3 input→output pairs + define exact structure
  │
  ├─ Task too large for single pass → Task Decomposition
  │   └─ Break into sequential stages; each feeds the next
  │
  ├─ No boundaries defined → Constraints & Guardrails
  │   └─ Define what NOT to do, length limits, format requirements
  │
  └─ Output structure undefined → Output Specification
      └─ Define headers, sections, length, style, tone explicitly

Core Process: The 4-D Framework

Apply this framework for every prompt optimization task.

1. DECONSTRUCT — Understand What's Really Being Asked

• What is the actual goal? (Often different from what's literally stated)
• What assumptions are unstated?
• What information is missing that the LLM will need?
• Who is the audience for the output?

2. DIAGNOSE — Identify What's Wrong or Missing

Score the prompt against these dimensions:

3. DEVELOP — Apply the Right Techniques

Select techniques based on diagnosis (see Technique Selection Decision Tree above). For detailed examples, see references/TECHNIQUES.md.

Role Assignment — Give the LLM a specific expert identity with credentials and methodology. Use when domain expertise matters. The more specific the role, the better the output quality.

Context Layering — Provide essential background in structured format: Background → Goal → Constraints → Output Format. Remove anything the LLM doesn't need.

Structured Reasoning — Ask the model to plan, check assumptions, and verify internally, then present a concise rationale or decision trace appropriate for the user. Avoid requesting hidden chain-of-thought verbatim; prefer "think privately, then summarize the key reasons."

Few-Shot Examples — Show 2-3 input→output pairs that demonstrate the pattern you want. This is the single most powerful technique for controlling output format and style.

Task Decomposition — Break complex tasks into sequential stages where each stage feeds the next. Prevents the LLM from trying to do everything at once and dropping quality.

Constraints & Guardrails — Define what NOT to do, set length limits, specify format requirements. LLMs perform better with clear boundaries than with open-ended freedom.

Output Specification — Define the exact structure, format, and content requirements of the output. Be explicit: headers, sections, length, style, tone.

4. DELIVER — Present the Optimized Prompt

• Show the complete optimized prompt in a code block or artifact
• Briefly explain key improvements (2-3 sentences, not a lecture)
• Note which techniques were applied and why
• If relevant, provide platform-specific tips (see references/PLATFORMS.md)
• Offer to iterate if the user wants refinements

Optimization Patterns

Pattern: Role + Context + Task + Format

The most common pattern. Works for 80% of prompt optimization needs.

You are a [SPECIFIC EXPERT] with expertise in [DOMAIN].

Context:
[ESSENTIAL BACKGROUND — 2-4 lines max]

Task:
[CLEAR, SPECIFIC OBJECTIVE]

Requirements:
- [CONSTRAINT 1]
- [CONSTRAINT 2]

Output format:
[EXACT STRUCTURE EXPECTED]

Pattern: Few-Shot + Structured Reasoning

Use for tasks requiring consistent format AND complex reasoning.

[ROLE AND CONTEXT]

Here are examples of the expected analysis:

Example 1:
Input: [SAMPLE]
Rationale: [CONCISE VISIBLE REASONS]
Output: [RESULT]

Example 2:
Input: [SAMPLE]
Rationale: [CONCISE VISIBLE REASONS]
Output: [RESULT]

Now analyze the following. Think privately through the reasoning, check for mistakes,
then provide the final output with a concise rationale.

Input: [ACTUAL TASK]

Pattern: Multi-Stage Pipeline

Use for complex tasks that benefit from decomposition.

Complete this analysis in three stages:

Stage 1 — Research:
[GATHER AND ORGANIZE INFORMATION]
Present findings as: [FORMAT]

Stage 2 — Analysis:
Using the research from Stage 1, [ANALYZE SPECIFIC ASPECTS]
Present analysis as: [FORMAT]

Stage 3 — Synthesis:
Based on your analysis, [PRODUCE FINAL DELIVERABLE]
Format: [FINAL OUTPUT SPECIFICATION]

Anti-Patterns to Fix

Context Enrichment

When the user's prompt references specific company data, projects, documents, or internal information, enrich the prompt with real context before optimizing.

When to enrich:

• User mentions "our company/team/project/product"
• References specific documents, emails, or meetings
• Uses domain terminology suggesting organizational context
• Mentions temporal markers like "Q3", "this sprint", "last month"

How to enrich: Use available tools (web search, connected integrations) to pull relevant context. Synthesize the key facts — goals, metrics, stakeholders, constraints, timelines — and inject them into the prompt's context section. Don't dump raw data — distill what the LLM actually needs to produce a useful output.

Evaluation Framework

When asked to evaluate a prompt (or when testing an optimized prompt), assess across these dimensions. See references/EVALUATION.md for the full methodology.

Quick Evaluation (Score 1-5 each)

A/B Testing Process

When comparing two prompt versions:

1. Define 3-5 test inputs covering typical, edge, and stress cases
2. Run both prompts against each test input
3. Blind-evaluate outputs (don't look at which prompt produced which)
4. Score on the relevant dimensions
5. Declare winner with reasoning

Platform-Specific Notes

See references/PLATFORMS.md for detailed platform optimization guides.

Reference Files

Load these as needed for deeper guidance: