Engineering Prompts

Engineering Prompts

---

LEVEL 1: QUICKSTART ⚡

Foundation First (Always Apply):

1. Clarity: Explicit instructions + success criteria
2. Context & Motivation: Explain WHY, not just WHAT
3. Positive Framing: Say what TO do, not what NOT to do

Then Assess:

1. Simple task? → Stop here. Foundation is enough.
2. Complex task? → Add advanced techniques sparingly
3. Test & Iterate: Measure effectiveness, refine based on results

Key insight: The best prompt achieves goals with minimum necessary structure.

---

LEVEL 2: DIAGNOSTICS 🔧

Output Problems → Solutions:

Process Problems → Fixes:

---

LEVEL 3: THE 12 TECHNIQUES 🛠️

Foundation (Always Apply)

Note: XML is less essential with Claude 4.x - use when genuinely needed.

Advanced (Apply When Needed)

---

LEVEL 4: DESIGN CHECKLIST 📋

Before Writing:

• Core task clear?
• Output format defined?
• Constraints known? (latency/​cost/​accuracy)
• One-off or repeated use?

Complexity Assessment:

• Simple (extraction, Q&A) → Foundation only
• Medium (analysis, code gen) → + CoT, examples
• Complex (research, novel problems) → + Chaining, role

Cost Optimization:

• Cache system prompts + reference docs (90% savings)
• Batch non-urgent work (50% savings)
• Skip CoT for simple tasks (saves 2-3x)

Deliverable:

• Prompt + techniques used + rationale + token estimate

---

LEVEL 5: ADVANCED TOPICS 🚀

Tool Integration

When to use MCP tools during prompt engineering:

Need latest practices?
└─ mcp__plugin_essentials_perplexity

Complex analysis needed?
└─ mcp__plugin_essentials_sequential-thinking

Need library docs?
└─ mcp__plugin_essentials_context7

Context Management

Prompt Caching:

• Cache: System prompts, reference docs, examples
• Savings: 90% on cached content
• Write: 25% of standard cost
• Read: 10% of standard cost

Long Context Tips:

• Place documents BEFORE queries
• Use XML tags: <document>, <source>
• Ground responses in quotes
• 30% better performance with proper structure

Token Optimization

Reducing Token Usage:

• Concise, clear instructions (no fluff)
• Reuse examples across calls (cache them)
• Structured output reduces back-and-forth
• Tool use instead of long context when possible

Cost-Specific Anti-Patterns

❌ Ignoring caching - Not leveraging repeated content (90% savings lost) ❌ Over-requesting CoT - Chain of thought for simple tasks (2-3x wasted) ❌ Redundant examples - 5 examples when 2 suffice ❌ No batching - Real-time calls for non-urgent work (50% savings lost)

See LEVEL 2: Diagnostics for general fixes.

---

LEVEL 6: REFERENCES 📚

Deep Dive Documentation

Detailed Technique Catalog:

• reference/​technique-catalog.md - Each technique explained with examples, token costs, combination strategies

Real-World Examples:

• reference/​examples.md - Before/​after pairs for coding, analysis, extraction, agent tasks

Research Papers:

• reference/​research.md - Latest Anthropic research, benchmarks, best practices evolution