SKILL.md
Engineering Prompts
LEVEL 1: QUICKSTART ⚡
Foundation First (Always Apply):
- Clarity: Explicit instructions + success criteria
- Context & Motivation: Explain WHY, not just WHAT
- Positive Framing: Say what TO do, not what NOT to do
Then Assess:
- Simple task? → Stop here. Foundation is enough.
- Complex task? → Add advanced techniques sparingly
- 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:
| Problem | Solution |
|---|---|
| Too generic | Add specificity + examples |
| Off-topic | Add context explaining goals |
| Wrong format | Examples or structured outputs |
| Unreliable on complex tasks | Prompt chaining |
| Unnecessary preambles | "Start directly with..." or structured outputs |
| Hallucinations | "If unsure, acknowledge uncertainty" |
| Shallow reasoning | Raise effort |
Process Problems → Fixes:
| Mistake | Fix |
|---|---|
| Over-engineering | Start minimal, add only what helps |
| Negative framing | Say what TO do instead |
| No motivation | Explain the WHY |
| No iteration | Test, measure, refine |
LEVEL 3: THE 12 TECHNIQUES 🛠️
Foundation (Always Apply)
| Technique | What It Does | Cost |
|---|---|---|
| 1. Clarity | Explicit instructions, success criteria | Minimal |
| 2. Context & Motivation | Explain WHY you want something | Minimal |
| 3. Positive Framing | Say what TO do, not what NOT to do | Minimal |
| 4. XML Structure | Separate sections in complex prompts | ~50-100 tokens |
Note: XML is less essential with Claude 4.x - use when genuinely needed.
Advanced (Apply When Needed)
| Technique | When to Use | Cost |
|---|---|---|
| 5. Effort / Thinking | Hard analysis, math, debugging | Higher effort = more tokens |
| 6. Prompt Chaining | Multi-step, complex tasks | Multiple calls |
| 7. Multishot Examples | Pattern learning, format | 200-1K each |
| 8. System Role | Domain expertise needed | Minimal |
| 9. Structured Outputs | Strict format requirements | Minimal |
| 10. Long Context | 20K+ token inputs | Better accuracy |
| 11. Context Budget | Repeated use, conversations | 90% savings |
| 12. Tool Docs | Function calling, agents | 100-500/tool |
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) → + higher effort, examples
- Complex (research, novel problems) → + Chaining, role
Cost Optimization:
- Cache system prompts + reference docs (90% savings)
- Batch non-urgent work (50% savings)
- Lower effort for simple tasks
Deliverable:
- Prompt + techniques used + rationale + token estimate
LEVEL 5: ADVANCED TOPICS 🚀
Tool Integration
When to use MCP tools during prompt engineering:
Need latest practices?
└─ WebSearch → Parallel Search
Need library docs?
└─ mcp__plugin_essentials_mcp (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 effort - High effort for simple tasks (tokens 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